betaFeedback

WorksComputer Science

GOFish: Full Field-of-View Gaussian Splatting and Mesh Extraction

Abstract

Robots, autonomous vehicles, and consumer 360° cameras use fisheye lenses for wide fields of view (FoVs). Simulation-ready digital twins require photorealistic rendering across the full FoV and an aligned collision mesh. However, prior fisheye Gaussian Splatting methods only rasterize up to 180° and do not produce meshes, while Gaussian surface reconstruction methods assume pinhole training images. We present GOFish, the first 3D Gaussian Splatting surface reconstruction method to train directly on fisheye images, rasterize FoVs beyond 180°, and extract an aligned mesh from the same scene. Calibrated densification recovers the surface across the entire FoV, while Calibrated Mip Filtering and Wide-Angle Appearance Modeling keep pixel blur and lens falloff out of the scene and mesh. Across the 200° FIORD and 205° FullCircle benchmarks, GOFish improves PSNR over the best prior method by 2.28 and 1.97 dB, respectively. It also outperforms fisheye rendering baselines within their own FoVs and achieves the highest mesh F-score on FIORD and the main ScanNet++ set. Together, the Gaussian scene and mesh form a LiDAR-free, simulation-ready digital twin that renders in real time, ray traces the full FoV, and provides collision geometry.

Introduction

Digital twins are virtual replicas of real environments used for robot simulation and training. To support visual perception and physical interaction, they must combine photorealistic appearance with aligned collision geometry. Fisheye cameras are commonly used for robotic perception and mapping [35] [21], while consumer 360∘360^\circ cameras pair back-to-back fisheye lenses for omnidirectional scene capture [5]. LiDAR scans are expensive to collect at scale, motivating recovery of both appearance and geometry directly from calibrated fisheye images over their full field of view (FoV).

No prior 3D Gaussian Splatting [17] method delivers both. Fisheye Gaussian Splatting rasterizers [3, 14, 22, 31, 32] cover at most 180∘180^\circ, while 3DGRT [24] ray traces the full FoV but bakes pixel blur into the scene; neither produces a mesh. Conversely, Gaussian Opacity Fields (GOF) [37] and other Gaussian surface methods [2, 7, 26, 27] extract aligned meshes but assume pinhole training images. Using fisheye captures therefore requires rectification, which discards peripheral rays and prevents rendering the native FoV required for accurate sensor simulation.

We present GOFish, the first 3D Gaussian Splatting surface reconstruction method to train directly on fisheye images, rasterize calibrated FoVs beyond 180∘180^\circ, and extract an aligned mesh from the same Gaussian scene. A calibration-defined domain and complete admission cover every valid ray, while calibrated densification recovers the surface across the FoV. We further introduce Calibrated Mip Filtering and Wide-Angle Appearance Modeling to keep view-specific sampling and camera response outside the shared scene representation used for meshing. Figure 1 illustrates GOFish’s full-FoV rendering and aligned mesh on a 200° FIORD capture [8], while Figure 6 demonstrates full-FoV ray tracing and mesh-based physics simulation.

Comparison of full-field-of-view Gaussian Splatting and mesh reconstruction results

Figure 1. GOFish reconstructs Hall_In from FIORD [8] as a Gaussian Splatting scene and aligned mesh over its full 200∘200^\circ field of view. 3DGUT + PPISP [31, 4] stops at 160∘160^\circ and does not extract a mesh. Figure 6 demonstrates full-FoV ray tracing and physical simulation.

Ray tracing and mesh-based physics simulation

Figure 6. Ray tracing and mesh-based physics simulation. 3DGUT ray traces its 160∘160^\circ FoV using 3DGRT [24] and does not produce a mesh, so simulated objects have no surface to collide with. GOFish covers the full 200∘200^\circ FoV and supports collisions.

GOFish leads every rendering metric across the 200° FIORD, 205° FullCircle, 127° ScanNet++, and pinhole Tanks & Temples benchmarks [5, 18, 34]. On average, it also outperforms all fisheye baselines within their respective fields of view, confirming higher rendering quality even without its coverage advantage. Furthermore, its meshes achieve the highest F-score on FIORD and the main ScanNet++ set. Beyond these gains, GOFish renders in real time, ray traces the full FoV, and provides collision geometry, turning native fisheye captures into LiDAR-free, simulation-ready digital twins.

In summary, our contributions are:

  1. Full Field-of-View Rasterization: Calibration-defined domains and complete admission extend Gaussian rasterization beyond 180° for the first time, with a 2.28 dB PSNR lead over the best prior baseline on FIORD.

  1. Wide-Angle Appearance Modeling: Camera-fixed attenuation captures steep rim falloff outside the 3D scene, improving FIORD PSNR by 3.17 dB over PPISP [4].

  1. Full Field-of-View Mesh Extraction: Calibrated densification and filtering enable an aligned full-FoV opacity-field mesh, achieving the highest mean F-score and recall among Gaussian surface methods on FIORD.

Related Work

Gaussian surface methods extract meshes from 3D Gaussian Splatting scenes [17]. Gaussian Opacity Fields (GOF) [37] constructs a 3D opacity field by compositing per-view Gaussian responses and taking the minimum across views. SOF [26] adds per-pixel sorting, geometry losses, and fast extraction, while CoMe [27] adds confidence-driven optimization. PGSR [2] uses planar Gaussians with geometric consistency. MiLo [7] jointly refines a Gaussian scene and its extracted mesh. Other approaches use surface-oriented primitives or explicit mesh connectivity [6, 12, 13, 39]. Existing Gaussian surface methods assume pinhole or rectified inputs, precluding direct mesh reconstruction from fisheye imagery.

Fisheye Gaussian Splatting methods train directly on distorted images. Fisheye-GS [22] and DirectFisheye-GS [32] modify the splatting projection, while Self-Calibrating GS [3] learns a lens deformation. 3DGUT [31] supports arbitrary camera models, and 3DGEER [14] evaluates Gaussians exactly along rays up to 180°. However, their rasterizers stop short of the full per-lens FoVs of the dual-fisheye 360° cameras used to capture FIORD and FullCircle [5, 8]. 3DGRT [24] ray traces the full field of view with one ray per pixel, baking pixel blur into the Gaussian scene while rendering substantially more slowly than rasterized methods. Omnidirectional methods [5, 19, 20] target panoramas and casual 360° captures. None of these methods includes mesh extraction. GOFish instead trains directly on fisheye images, rasterizes the full field of view, and extracts a mesh from the same Gaussian scene.

When unmodeled, sampling and appearance effects are absorbed into the Gaussian scene, degrading rendering and mesh extraction. Mip-Splatting [36] suppresses aliasing in pinhole Gaussian Splatting with 3D and screen-space filters. Gaussian surface methods use VastGaussian’s appearance model [23] to keep photometric inconsistencies from creating spurious geometry [26, 37]. PPISP [4] instead limits corrections to camera effects, but its low-order vignetting misses steep rim falloff in wide-FoV fisheye images. GOFish adapts Mip-Splatting’s filters to calibrated fisheye cameras and extends PPISP to model this falloff.

Background

3D Gaussian Splatting (3DGS) [17] represents a scene using Gaussians, each with mean μi\mu_i, covariance Σi\Sigma_i, opacity oio_i, and view-dependent color. Its rasterizer linearizes a pinhole projection at each Gaussian mean [40], so the standard rasterizer cannot consume native fisheye images or represent rays beyond 90°. GOFish instead builds on 3DGUT [31], which evaluates Gaussians directly in 3D along calibrated pixel rays. A ray is r(t)=o+tdr(t) = o + td, where oo is the camera center, dd a unit direction, and t≥0t \ge0 the distance along the ray. For Gaussian ii, 3DGUT’s ray-maximum response occurs at the minimum Mahalanobis distance over the valid ray interval:

δi=min⁡t∈I∥r(t)−μi∥Σi−1,αi=min⁡(αmax⁡,oie−δi2/2).(1)\delta_i = \min_{t \in\mathcal{I}} \lVert r(t) - \mu_i \rVert_{\Sigma_i^{-1}}, \qquad\alpha_i = \min\left(\alpha_{\max}, o_i e^{-\delta_i^2/2}\right). \tag*{(1)}

Here I\mathcal{I} is the valid distance interval, and αmax⁡\alpha_{\max} caps the response. GOF [37] meshes this response, while 3DGRT [24] ray traces the same Gaussians.

3DGUT depth-sorts Gaussians and composites their alphas as in 3DGS. However, for a ray at angle θ\theta with optical-axis component dzd_z, camera-space depth z=tdzz = t d_z becomes negative beyond 90°, so sorting by depth reverses the near-to-far order. Its tile selection also approximates each Gaussian’s image extent by projecting unscented-transform sigma points [15] through the camera. Because αi\alpha_i instead depends on the ray-to-Gaussian Mahalanobis distance δi\delta_i, this bound can omit tiles whose rays reach the Gaussian. Figure 2 illustrates the tile mismatch and two field-of-view limits: the half-angle crop stops at 80° from the axis, and the positive-depth test rejects valid content beyond 90°.

Comparison of wide-angle rasterization failures and GOFish rendering

Figure 2. 3DGUT failures over a full field of view. (a) Its half-angle limit cuts rays beyond 80° from the axis (orange), and its positive-depth test fails beyond 90° (red). (b) Tiles come from the projected sigma-point extent (blue) but alpha from the exact ray response (red), so the Gaussian is skipped in the hatched tiles. (Bottom) 3DGUT + PPISP leaves a cutoff ring and tile-shaped smears, while GOFish renders the full FoV cleanly.

Background

Fisheye Gaussian Splatting methods train directly on distorted images. Fisheye-GS [22] and DirectFisheye-GS [32] modify the splatting projection, while Self-Calibrating GS [3] learns a lens deformation. 3DGUT [31] supports arbitrary camera models, and 3DGEER [14] evaluates Gaussians exactly along rays up to 180°. However, their rasterizers stop short of the full per-lens FoVs of the dual-fisheye 360° cameras used to capture FIORD and FullCircle [5, 8]. 3DGRT [24] ray traces the full field of view with one ray per pixel, baking pixel blur into the Gaussian scene while rendering substantially more slowly than rasterized methods. Omnidirectional methods [5, 19, 20] target panoramas and casual 360° captures. None of these methods includes mesh extraction. GOFish instead trains directly on fisheye images, rasterizes the full field of view, and extracts a mesh from the same Gaussian scene.

When unmodeled, sampling and appearance effects are absorbed into the Gaussian scene, degrading rendering and mesh extraction. Mip-Splatting [36] suppresses aliasing in pinhole Gaussian Splatting with 3D and screen-space filters. Gaussian surface methods use VastGaussian’s appearance model [23] to keep photometric inconsistencies from creating spurious geometry [26, 37]. PPISP [4] instead limits corrections to camera effects, but its low-order vignetting misses steep rim falloff in wide-FoV fisheye images. GOFish adapts Mip-Splatting’s filters to calibrated fisheye cameras and extends PPISP to model this falloff.

Background

3D Gaussian Splatting (3DGS) [17] represents a scene using Gaussians, each with mean μi\mu_i, covariance Σi\Sigma_i, opacity oio_i, and view-dependent color. Its rasterizer linearizes a pinhole projection at each Gaussian mean [40], so the standard rasterizer cannot consume native fisheye images or represent rays beyond 90°. GOFish instead builds on 3DGUT [31], which evaluates Gaussians directly in 3D along calibrated pixel rays. A ray is r(t)=o+tdr(t) = o + td, where oo is the camera center, dd a unit direction, and t≥0t \ge0 the distance along the ray. For Gaussian ii, 3DGUT’s ray-maximum response occurs at the minimum Mahalanobis distance over the valid ray interval:

δi=min⁡t∈I∥r(t)−μi∥Σi−1,αi=min⁡(αmax⁡,oie−δi2/2).(2)\delta_i = \min_{t \in\mathcal{I}} \lVert r(t) - \mu_i \rVert_{\Sigma_i^{-1}}, \qquad\alpha_i = \min\left(\alpha_{\max}, o_i e^{-\delta_i^2/2}\right). \tag*{(2)}

Here I\mathcal{I} is the valid distance interval, and αmax⁡\alpha_{\max} caps the response. GOF [37] meshes this response, while 3DGRT [24] ray traces the same Gaussians.

3DGUT depth-sorts Gaussians and composites their alphas as in 3DGS. However, for a ray at angle θ\theta with optical-axis component dzd_z, camera-space depth z=tdzz = t d_z becomes negative beyond 90°, so sorting by depth reverses the near-to-far order. Its tile selection also approximates each Gaussian’s image extent by projecting unscented-transform sigma points [15] through the camera. Because αi\alpha_i instead depends on the ray-to-Gaussian Mahalanobis distance δi\delta_i, this bound can omit tiles whose rays reach the Gaussian. Figure 2 illustrates the tile mismatch and two field-of-view limits: the half-angle crop stops at 80° from the axis, and the positive-depth test rejects valid content beyond 90°.

Method

GOFish builds on 3DGUT [31] to train one Gaussian scene directly from fisheye images over the camera’s full field of view (FoV) and extract an aligned mesh. Calibration defines valid pixels, complete admission assigns Gaussians to tiles, and calibrated densification recovers the surface. Calibrated Mip Filtering and Wide-Angle Appearance Modeling keep view-specific blur and sensor falloff out of the scene and mesh. GOFish then extracts the mesh from the resulting pixel-filter-free opacity field. Together, the Gaussian scene and mesh form a simulation-ready digital twin.

Full Field-of-View Rasterization

3DGUT’s rasterizer defines a ray for each pixel, assigns Gaussians to image tiles, and composites them along each ray in depth order. As Figure 2 shows, each step breaks at wide angles: rays outside its 160° field of view are cut and receive no supervision, projected extents omit Gaussians from tiles their rays reach, and depth order reverses beyond 90°. GOFish replaces these steps with a calibration-defined domain, complete admission, and ray-distance sorting, extending rasterization to the full calibrated FoV.

A pixel is valid if it lies in the calibrated image domain and defines a unique ray. For a radial fisheye mapping ray angle θ\theta to image radius ρ(θ)\rho(\theta), rays are unique up to

θmax⁡=sup⁡{a∈(0,π]:ρ′(θ)>0 ∀θ∈[0,a]},(3)\theta_{\max} = \sup\{a \in(0,\pi] : \rho'(\theta) > 0\ \forall\theta\in[0,a]\}, \tag*{(3)}

where ρ′=dρ/dθ\rho' = d\rho/d\theta. The calibration itself, rather than a preset cutoff, therefore bounds the valid domain of any radial central camera, including FoVs wider than 200°. Rendering, admission, and meshing share this domain, which Appendix § 12.§ 1 defines formally.

Let Dv\mathcal{D}_v be the valid image domain of view vv and Pv\mathcal{P}_v its pixel centers. We require complete admission: Gaussian ii is drawn in every tile T\mathcal{T} containing a valid pixel whose ray intersects its effective support Si\mathcal{S}_i, the region where the renderer can return a visible alpha:

min⁡u∈T∩Pv∩Dvδi(u)≤ri⟹i is admitted to T.(4)\min_{u\in\mathcal{T}\cap\mathcal{P}_v\cap\mathcal{D}_v}\delta_i(u) \le r_i \quad\Longrightarrow\quad i\text{ is admitted to }\mathcal{T}. \tag*{(4)}

Here δi(u)\delta_i(u) is pixel uu’s ray-to-Gaussian Mahalanobis distance from Eq. (1). We cap the opacity-dependent radius rir_i at radm=3r_{\mathrm{adm}}=3, conservatively enclosing 3DGUT’s renderer cutoff. 3DGUT’s projected sigma-point extents can omit tiles whose rays reach the Gaussian, and conventional projection-based bounds fail beyond 90°. We instead use three conservative rejection tests, defined in Appendix § 12.§ 2, each of which drops a Gaussian only if it proves that every valid ray in the tile misses Si\mathcal{S}_i. Together, they guarantee:

Proposition 1 (Complete admission). Given a central camera, exact arithmetic, and conservative enclosures, the three tests of Appendix § 12.§ 2 admit Gaussian ii to every tile containing a valid pixel whose ray intersects Si\mathcal{S}_i.

The same appendix gives the proof. In practice, we round every rejection bound outward to keep the tests conservative under floating-point arithmetic. Because dz<0d_z<0 beyond 90°, depth z=tdzz=td_z reverses the near-to-far order. We therefore sort the admitted Gaussians at each pixel by ray distance, following StopThePop [25].

Calibrated Mip Filtering

Complete ray coverage alone is not enough: each image pixel averages radiance over a finite footprint whose size and shape vary across a fisheye image. The 3D Gaussian scene is shared by every view and later meshed, so this view-specific blur must remain outside it. Without a pixel filter, however, the exact ray response can match the blur only by baking it into the scene. For pinhole cameras, Mip-Splatting [36] separates a view-independent 3D scene filter from a 2D pixel filter. The 3D filter makes scene detail to what the finest training view can resolve, while the 2D filter matches each view’s pixel footprint.

As illustrated in Figure 3, we adapt this 3D filter to calibrated cameras by deriving its scale from the camera Jacobian while leaving opacity unchanged. For each training view vv observing Gaussian ii, the calibrated Jacobian Ji,v\mathbf{J}_{i,v} be the 2×32\times3 Jacobian of vv’s lens-distorted projection at μi\boldsymbol{\mu}_i. The smaller singular value σmin⁡(Ji,v)\sigma_{\min}(\mathbf{J}_{i,v}) is the coarser of the view’s two image-space sampling rates. Using Mip-Splatting’s 3D-filter variance of 0.2 px20.2\ \mathrm{px}^2, we set

Calibrated Mip Filtering diagram

Figure 3. Calibrated Mip Filtering. The finest training view’s Jacobian sets the 3D filter +ϵi2I3+\epsilon_i^2 I_3 in the stored Gaussian covariance Σi\Sigma_i (blue). Rendering view vv adds the footprint +γJ+J+T+\gamma J^+J^{+\mathsf{T}} and rescales opacity by mm (orange), while meshing uses Σi\Sigma_i directly.

ϵi=0.2max⁡vσmin⁡(Ji,v).\epsilon_i = \frac{\sqrt{0.2}}{\max_v \sigma_{\min}(\mathbf{J}_{i,v})}.

The maximum selects the finest observing view, and dividing by its sampling rate converts the pixel-space filter width to a world-space radius. From here on, Σi\Sigma_i denotes Gaussian ii’s covariance after adding ϵi2I3\epsilon_i^2 I_3.

The finest view sets the 3D filter, so a Gaussian may project to less than one pixel in a coarser view. For each view vv, we therefore add the minimum-trace 3D covariance that projects to 3DGS’s fixed screen-space filter of variance γ=0.3 px2\gamma=0.3\ \mathrm{px}^2[17]. We also rescale opacity to preserve the image integral. Let J=Ji,vJ=J_{i,v} have full row rank, and let J+J^+ denote its pseudo-inverse. Then

C=Σi+γJ+J+T,m=det⁡(JΣiJT)det⁡(JCJT).(5)C=\Sigma_i+\gamma J^+J^{+\mathsf{T}},\qquad m=\sqrt{\frac{\det(J\Sigma_iJ^{\mathsf{T}})}{\det(JCJ^{\mathsf{T}})}}. \tag*{(5)}

View vv then evaluates Eq. (1) with covariance CC and opacity m oim\ o_i. CC and mm are recomputed for every view and never stored, so neither enters mesh extraction. The covariance lift has the following minimality guarantee:

Proposition 2 (Minimum-trace footprint lift). For full-row-rank JJ and γ>0\gamma>0, γJ+J+T\gamma J^+J^{+\mathsf{T}} is the unique positive-semidefinite matrix XX of minimum trace satisfying JXJT=γI2JXJ^{\mathsf{T}}=\gamma I_2. It adds no variance along the viewing-ray direction, which spans the null space of JJ.

The added covariance projects through JJ to γI2\gamma I_2, matching the form of Mip-Splatting’s 2D filter but using 3DGS’s variance. Within the same linearization, the ray-maximum response projects to a 2D Gaussian with covariance JCJTJCJ^{\mathsf{T}}, whose image integral scales with det⁡(JCJT)1/2\det(JCJ^{\mathsf{T}})^{1/2}. The factor mm in Eq. (5) cancels this growth, so the footprint preserves the image integral whenever the αmax⁡\alpha_{\max} cap is inactive. Mip-Splatting instead rescales opacity by the 3D volume ratio det⁡Σi/det⁡C\sqrt{\det\Sigma_i/\det C}, which preserves the Gaussian’s 3D integral. Proposition 3 shows that this ratio equals mm times a ray-dependent factor of at most one:

Proposition 3 (Ray-maximum normalization). After linearizing the camera projection at μi\mu_i, let dd denote the viewing-ray direction. The conditional variance along this ray is s(Σ)=1/(dTΣ−1d)s(\Sigma)=1/(d^{\mathsf{T}}\Sigma^{-1}d). The 3D volume ratio and image-space normalization are then related by

det⁡Σidet⁡C=ms(Σi)s(C)≤m,(6)\sqrt{\frac{\det\Sigma_i}{\det C}}=m\sqrt{\frac{s(\Sigma_i)}{s(C)}}\leq m, \tag*{(6)}

with equality exactly when dd is an eigenvector of Σi\Sigma_i.

Using it would therefore dim Gaussians whose principal axes are not aligned with the ray, darkening surfaces viewed at grazing angles. Appendix A.3 proves both propositions.

Calibrated Densification

The calibrated Jacobian also drives the densification that recovers the surface. A 3D gradient measures image-space displacement in world units, whose pixel equivalent varies with distance and ray angle, while a pinhole projection fails beyond 90°. We therefore map each mean’s gradient into the calibrated image with J+TJ^{+\mathsf{T}}, allowing one image-space threshold and SOF’s densification schedule to serve the full FoV. Appendix A.4 details the mapping and schedule. Calibrated Mip Filtering prevents Gaussians from fitting each view’s blur; Section 6.2 shows that removing it produces several times more Gaussians without improving quality.

Wide-Angle Appearance Modeling

Exposure, vignetting, and rim falloff are image-formation effects that must remain outside the 3D scene. Gaussian surface methods handle photometric variation with VastGaussian’s per-image appearance network [23, 26, 37], but its corrections are not limited to camera effects. We instead build on PPISP [4], which restricts corrections to exposure, vignetting, color, and camera response. At inference, its controller predicts exposure and color for novel views in real time. CoMe [27] reports that PPISP outperforms its improved VastGaussian model and nearly matches the full method on pinhole scenes. However, PPISP’s low-order vignetting does not capture the steep rim falloff, leaving the darkest rim pixels too bright and their neighbors too dark. This residual is baked into Gaussian colors, making Gaussians seen at the rim too dark near the center of other views.

We therefore multiply PPISP’s vignetting VV by a per-camera wide-angle attenuation WW. At pixel uu of view vv,

I^(u)=CRF⁡(Hv(VWevI(u))),W(u)=w(ρ)e−a(u)(7)\hat{I}(u)=\operatorname{CRF}\left(H_v\left(VW e_v I(u)\right)\right), \qquad W(u)=w(\rho)e^{-a(u)} \tag*{(7)}

where II is the rendered radiance, eve_v is the view’s exposure, HvH_v is its color transform, and CRF is the camera response function. The normalized radius ρ\rho is the distance of uu from the principal point divided by the larger image dimension.

The radial profile ww is nonincreasing, and a≥0a \ge0 is a low-resolution flat-field log-attenuation map. Both yield unit attenuation over the inner half of the image radius: w=1w=1 and a=0a=0. This anchoring prevents either from absorbing global exposure changes. We optimize both terms as detailed in Appendix A.4. WW is applied after Gaussian rendering and shared across a camera’s views, so it models only camera-fixed effects and never enters the 3D scene or mesh. For pinhole cameras, setting W=1W=1 reduces the model to PPISP.

Full Field-of-View Mesh Extraction

GOFish uses the same ray-maximum response as GOF [37], enabling opacity-field mesh extraction over every training view’s full calibrated domain. Let αi,v(x)\alpha_{i,v}(x) be Gaussian ii’s alpha from Eq. (1) on the ray segment from view vv to point xx, evaluated with stored covariance Σi\Sigma_i and opacity σi\sigma_i. We omit the per-view pixel filter to keep view-specific footprints out of the field and mesh. The opacity field is

O(x)=min⁡v[1−∏i(1−αi,v(x))](8)O(x)=\min_v\left[1-\prod_i\left(1-\alpha_{i,v}(x)\right)\right] \tag*{(8)}

where vv ranges over training views with a valid ray to xx, and ii over Gaussians admitted along that ray by the mesh rule of Appendix A.5. Following GOF, points that no training view observes remain occupied, with O(x)=1O(x)=1. A cropped renderer therefore leaves points visible only through peripheral rays filled in, whereas full-FoV coverage evaluates them. We extract the O(x)=0.5O(x)=0.5 level set using the tetrahedral opacity-field readout and filters inherited from SOF and GOF; Appendix A.5 details both steps.

GOFish optionally adds SOF’s Geometry Objective [26], replacing pinhole depth with distance along calibrated rays. Its opacity term evaluates the 3D scene without the per-view pixel footprint, so it supervises the same footprint-free per-view opacity from which the mesh field OO is built. Appendix A.4 gives the adapted objective.

Experiments

GOFish extends the 3DGUT [31] codebase. We evaluate GOFish with and without SOF’s Geometry Objective [26] from § 6.5. Each variant uses one parameter set across datasets except ScanNet++, where, following prior surface methods, we add CoMe’s confidence and normal-variance losses [27] and use SOF’s scene-extent rule for large-Gaussian pruning. Appendices A.4 and B provide additional optimization and evaluation details.

Datasets. FIORD [8] and FullCircle [5] use dual-fisheye 360° cameras. FIORD pairs synchronized nominal 200° views with LiDAR; FullCircle provides roughly 205° views without geometry ground truth. ScanNet++ [34] provides 127° fisheye DSLR images and laser-scan ground truth. We use 3DGUT’s six scenes, while Appendix C additionally evaluates CoMe’s set. For pinhole evaluation, we use six Tanks & Temples [18] scenes with ground-truth geometry. Appendix B.1 covers FIORD’s registration and calibration issues; Appendix B.2 details scenes, resolutions, and splits.

Baselines. We use each baseline’s released code and recipe. Fisheye baselines retain their native camera pipelines, including 3DGUT’s angular crop, while surface baselines train on the rectifications marked in the tables. Appendix § 13.3 details adaptations and hardware exceptions.

Metrics. We report PSNR, SSIM, and LPIPS on held-out test views. All tabulated GOFish rendering scores use the complete-admission renderer of § 6.1, while FPS uses the real-time renderer detailed in Appendix § 12.§ 8. Full scores each camera’s calibrated FoV, and unrendered pixels count as black. Common (90∘90^\circ) scores every method on the shared rectified FIORD images used by the surface baselines. For geometry, we report precision, recall, and F-score under the protocols detailed in Appendix § 13.§ 2: a 10 cm threshold on FIORD, CoMe’s 5 cm protocol on ScanNet++, and the official evaluator on Tanks & Temples.

Results

Across its base and Geometry Objective variants, GOFish ranks first in every full-FoV rendering metric on the three fisheye benchmarks in Table 1. Its PSNR exceeds the best prior result by 2.28 dB on FIORD, 1.97 dB on FullCircle, and 0.46 dB on ScanNet++ [5, 8, 34]. The gains are not merely due to coverage; on average, GOFish also outper-

MethodFoVFIORD [8] Full (200°) PSNR↑SSIM↑LPIPS↓GOFish @ FoV ΔPSNR↑ΔSSIM↑ΔLPIPS↑FullCircle [5] Full (205°) PSNR↑SSIM↑LPIPS↓GOFish @ FoV ΔPSNR↑ΔSSIM↑ΔLPIPS↑ScanNet++ [34] Full (127°) PSNR↑SSIM↑LPIPS↓
Fisheye-GS* [22]145°17.780.6850.395+5.06+0.119+0.16714.010.6300.495+14.05+0.263+0.39726.280.8950.256
3DGUT [31]160°19.860.7400.316+3.95+0.057+0.04621.580.8420.262+1.62+0.013+0.03028.930.9130.214
3DGUT + PPISP [31, 4]160°20.180.7430.312+4.04+0.058+0.04821.450.8410.262+1.88+0.014+0.03028.930.9130.213
3DGEER* [14]178°12.790.6790.328+2.03+0.026+0.02712.650.7150.336+4.25+0.040+0.03628.590.9140.201
Self-Calibrating GS* [3]180°17.730.6880.352+5.65+0.129+0.13517.290.6690.435+10.44+0.219+0.29127.670.8850.240
DirectFisheye-GS [32]180°22.080.7530.304+2.21+0.070+0.07920.670.8050.292+5.31+0.053+0.06929.280.9100.207
3DGRT [24]Full15.840.6850.343+8.52+0.142+0.13323.320.8700.228+2.15+0.005+0.02028.830.9040.278
3DGRT + PPISP [24, 4]Full18.420.7120.313+5.94+0.115+0.10323.500.8720.222+1.97+0.004+0.01529.710.9110.247
GOFish (Ours)Full24.360.8270.21025.470.8760.20730.170.9210.172
+ Geometry ObjectiveFull24.180.8200.21725.310.8710.21330.120.9210.171

Table 1. Quantitative comparison against wide-angle rendering baselines. The best, second best, and third best entries are highlighted. GOFish’s gain at each baseline’s FoV is highlighted. * Appendix B.3 details baseline adaptations for wide-angle rendering.

forms all released fisheye renderers within their respective fields of view. Figure 4 shows that prior methods blur, smear, or ring near the FoV boundary, whereas GOFish keeps walls, windows, and branches sharp and correctly exposed to the rim. Figure 6 shows the recovered periphery in full-FoV reflections and the extracted mesh supporting collision physics. Appendix § 14.§ 7 shows additional scenes, and Appendix § 14.§ 6 reports per-scene metrics.

Qualitative rendering results comparing fisheye rendering methods on four scenes

Figure 4. Qualitative rendering results on FIORD [8] and FullCircle [5]. Held-out views span the full FoV; metrics are per-scene.

GOFish leads both full-FoV rendering and FIORD mesh F-score in Table 2. Appendix § 14.§ 1 confirms the lead across three runs, with base GOFish at 0.245±0.001 and PGSR [2] at 0.239 ± 0.002. The Geometry Objective [26] raises F-score further to 0.249. Calibrated densification recovers more surface across the FoV, giving GOFish the highest recall and F-score. Appendix § 14.§ 1 shows that GOFish leads within the shared 90∘90^\circ field of view and for rays beyond 90∘90^\circ from the optical axis, so wider coverage alone does not explain the advantage. Figure 5 shows broad peripheral patches or holes in surface-baseline meshes, while GOFish preserves finer structure.

MethodFoVFIORD [8]ScanNet++ [34]
Full (200°)Common (90°)GeometryFull (127°)Geometry
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓P ↑R ↑F ↑PSNR↑SSIM↑LPIPS↓P ↑R ↑F ↑
GOF [37]Rectified17.030.6760.35328.200.8830.2780.1680.2480.19827.540.9040.2120.5790.6290.601
PGSR [2]Rectified16.810.6700.36328.440.8920.2730.2250.2930.23927.390.8980.2410.6980.6770.687
SOF [26]Rectified16.620.6600.36326.150.8520.3110.1630.2400.19227.100.8980.2180.6200.6670.641
MiLo [7]Rectified17.050.6800.36828.210.8820.2970.1800.2610.21027.610.9030.2340.6510.7010.675
CoMe [27]Rectified17.640.6960.33227.430.8810.2720.1900.2960.22728.110.9090.1900.6520.7300.687
GOFish (Ours)Full24.360.8270.21029.000.9110.2210.2000.3310.24530.170.9210.1720.6370.7950.706
+ Geometry ObjectiveFull24.180.8200.21728.510.9060.2290.2020.3400.24930.120.9210.1710.6300.8010.704

Table 2. Quantitative comparison against surface reconstruction baselines. The best, second best, and third best entries are highlighted. Full scores each calibrated FoV and Common the shared rectified images; § 7 defines both.

Qualitative mesh results on FIORD and FullCircle

Figure 5. Qualitative mesh results on FIORD [8] and FullCircle [5]. GOFish uses the Geometry Objective. Metrics are per-scene.

Across its two variants, GOFish leads rendering, recall, and F-score on the main ScanNet++ set. Appendix § 14.§ 2 evaluates the ScanNet++ scene set used by CoMe [27] and provides the full pinhole comparison. On Tanks & Temples [18], GOFish renders best, and Appendix § 14.§ 7 shows visibly fewer rendering artifacts than the surface baselines. A single Gaussian scene thus provides both the best full-FoV rendering and, on FIORD and the main ScanNet++ set, the most complete reconstructed surface.

Rendering Gain by Ray Angle

Figure 7 compares GOFish with 3DGUT + PPISP [4, 31] across ray angles and isolates the contribution of Wide-Angle Appearance Modeling. Within the baseline’s 160∘160^\circ FoV, GOFish gains 4.04.0 dB on FIORD and 1.91.9 dB on FullCircle, of which appearance modeling accounts for only 0.40.4 and 0.20.2 dB, respectively. Complete admission, calibrated densification, and Calibrated Mip Filtering therefore drive most of the shared-region improvement. The FIORD gain peaks at 5.65.6 dB in the 75∘75^\circ–90∘90^\circ bin spanning the baseline’s 80∘80^\circ cutoff, while FullCircle reaches 6.96.9 dB beyond 100∘100^\circ, where the baseline renders nothing. In the outer-most bin, Wide-Angle Appearance Modeling adds 7.97.9 dB on FIORD and 1.11.1 dB on FullCircle. On FIORD, this exceeds GOFish’s lead over the baseline—without appearance modeling, GOFish renders the dark rim too brightly and scores below the baseline’s unrendered black pixels.

Figure 7. Rendering gain by ray angle on FIORD [8] (cyan) and FullCircle [5] (magenta). PSNR gain of GOFish over 3DGUT + PPISP [31, 4] (solid) and over GOFish without Wide-Angle Appearance (dashed). As in Figure 2, 3DGUT + PPISP cuts rays beyond 80° (orange) and fails beyond 90° (red).

Component Analysis

Table 3 cumulatively adds GOFish’s components to 3DGUT + PPISP. Full Field-of-View Rasterization improves rendering but leaves F-score at 0.1390.139, confirming that coverage alone does not recover the surface. Adding calibrated densification and Calibrated Mip Filtering then raises F-score to 0.2400.240 while PSNR remains nearly unchanged. Together, these stages account for 3.6 dB of GOFish’s 4.0 dB gain within the baseline’s 160° view. Wide-Angle Appearance Modeling raises Full PSNR by 3.17 dB to 24.36 dB, consistent with modeling rim falloff as an image-formation effect outside the Gaussian scene. Finally, the optional Geometry Objective raises F-score to 0.249 at a cost of 0.18 dB PSNR. The reference rows show that neither 3DGRT’s full-FoV ray tracing [24] nor adding SOF’s losses [26] to 3DGUT yields a competitive mesh.

VariantFoVPSNR↑SSIM↑LPIPS↓F↑Train↓FPS↑
GOF [37]90°17.030.6760.3530.198104m11.0
3DGUT + PPISP [24, 4]Full18.420.7120.3130.159108m22.3
3DGUT + SOF [26, 31]160°18.460.6710.3740.168111m17.9
3DGUT + PPISP [31, 4]160°20.180.7430.3120.14129m84.3
+ Full-FoV RasterizationFull21.210.7980.2530.139162m91.1
+ Calibrated Mip FilteringFull21.190.8040.2390.240103m81.1
+ Wide-Angle AppearanceFull24.360.8270.2100.245107m79.3
+ Geometry ObjectiveFull24.180.8200.2170.249116m86.8

Table 3. Component analysis on FIORD [8]. Components are added cumulatively. + Calibrated Mip Filtering includes calibrated densification, and + Wide-Angle Appearance completes base GOFish. Appendix § 13.§ 2 details the configurations.

The split ablation in Appendix § 14.§ 1 confirms the division of labor predicted in § 6.2. No individual part—the image-space statistic, SOF’s schedule, or the filter—is sufficient alone. For the two scenes with filtered and unfiltered meshes, F-score is essentially unchanged, indicating that calibrated densification drives surface recovery. At matched rendering quality, Calibrated Mip Filtering makes that recovery tractable by reducing the Gaussian count by 2.2–5.4× and preventing out-of-memory failures during training and meshing on a 24 GB GPU.

On an NVIDIA A10G GPU, the compact scene and real-time renderer reach 79 FPS on FIORD, nearly matching 3DGUT + PPISP at 84 FPS and far exceeding 3DGRT + PPISP at 22 FPS. Appendix § 12.§ 8 shows that the real-time renderer trails the complete-admission renderer by only 0.03 dB in Full PSNR. Together, the scene and mesh form a simulation-ready digital twin with full-FoV rendering quality, real-time speed, ray tracing, and collision geometry.

Limitations and Future Work

GOFish inherits opacity-field limitations [26, 37]. Taking the minimum across views can misplace reflective surfaces or carve thin structures, while sparse observations can leave holes and mesh overfill, as illustrated in Appendix § 14.3. Full-FoV coverage cannot recover information lost to sparse sampling or rim attenuation, and reconstruction remains sensitive to calibration and pose errors documented in Appendix § 13.§ 1. Among our datasets, only FIORD provides geometry ground truth beyond 180°. GOFish assumes static scenes and central camera models, so moving content, rolling shutter, and non-central cameras remain unsupported. Complete admission and per-pixel sorting slow training relative to rendering-only 3DGUT, although Appendix § 14.§ 2 shows that GOFish remains faster than GOF end to end on FIORD and renders in real time.

Depth supervision, normal priors, or Taubin smoothing [28, 30] could reduce surface artifacts. Sparse optimization and pruning [9, 10, 29] could reduce training cost and improve rendering speed. Selective super-resolution [1] could improve the undersampled rim, and extending complete admission to rolling-shutter and non-central cameras would broaden camera support. Agentic 3D modeling pipelines [11] could build on GOFish digital twins by refining their meshes, simplifying colliders, and adding semantic labels to create task-ready simulation environments for training robot policies.

Conclusion

We introduced GOFish, the first 3D Gaussian Splatting surface reconstruction method to train directly on fisheye images, rasterize fields of view (FoVs) beyond 180°, and extract an aligned mesh from the same Gaussian scene. A calibration-defined domain and complete admission cover every valid ray, while calibrated densification recovers the surface across the FoV. Calibrated Mip Filtering and Wide-Angle Appearance Modeling keep view-specific sampling and camera effects out of the Gaussian scene and mesh. GOFish leads every rendering metric across the 200° FIORD, 205° FullCircle, 127° ScanNet++, and pinhole Tanks & Temples benchmarks. Its meshes also achieve the highest F-score on FIORD and the main ScanNet++ set. Together, the Gaussian scene and mesh form a LiDAR-free, simulation-ready digital twin that renders in real time, ray traces the full FoV, and provides collision geometry.

Acknowledgements

This work was completed during an internship at Amazon. We thank Ziyang Long, Mellon M. Zhang, Chenqi Li, Sarosij Bose, Srikumar Sastry, Amin Kheradmand, Yongqing Ye, David Stafford, Huixian Gong, Yun He, Zeping Liu, Haiyang Ying, Alex Hanson, Chengyue Huang, and Chen Xu for helpful discussions and feedback on this work.

Additional Method Details

We make the view index explicit, so Ji,vJ_{i,v}, Ci,vC_{i,v}, and mi,vm_{i,v} denote the quantities of Eq. (5) for view vv. Throughout, ∥⋅∥\| \cdot\| is the Euclidean norm for vectors and spectral norm for matrices, ∥⋅∥F\| \cdot\|_F is the Frobenius norm, and B(c,r)B(c,r) is the closed Euclidean ball with center cc and radius rr.

Camera Domain

The camera domain determines which pixels participate in rendering, supervision, admission, and meshing. For view vv, let Uv⊂R2U_v \subset\mathbb{R}^2 be the continuous calibrated image domain and PvP_v its finite set of raster pixel centers. Let dv(u)d_v(u) denote the unit ray direction obtained by unprojecting uu, when that direction is unique. We define the valid domain as Dv={u∈Uv:dv(u) exists uniquely}D_v = \{u \in U_v : d_v(u)\text{ exists uniquely}\}, and rendering and admission operate on Pv∩DvP_v \cap D_v. For the OpenCV fisheye model, a ray at angle θ\theta from the optical axis maps to image radius ρ(θ)=θ(1+k1θ2+k2θ4+k3θ6+k4θ8)\rho(\theta) = \theta(1 + k_1\theta^2 + k_2\theta^4 + k_3\theta^6 + k_4\theta^8). Here k1,…,k4k_1,\ldots,k_4 are the calibrated radial-distortion coefficients. Rays are unique up to θmax⁡\theta_{\max} of Eq. (3). Dataset masks do not alter DvD_v. Masks apply only to supervision, densification statistics, and PPISP [4] pooling. Rendering, reporting, and meshing use the full domain. Losses that compare neighboring pixels, such as SSIM, include only pixels whose full stencil lies within the domain. We measure camera-local distance as view-space depth for pinhole cameras and radial ray distance for fisheye cameras, and use the same convention for SOF’s distortion coordinate [26].

Complete Admission

Let Σi=BiBiT\Sigma_i = B_iB_i^\mathsf{T} be the covariance used by the renderer and define the corresponding Mahalanobis distance as qi(x)2=(x−μi)TΣi−1(x−μi)q_i(x)^2 = (x-\mu_i)^\mathsf{T}\Sigma_i^{-1}(x-\mu_i). The closed effective support Si={x:qi(x)≤ri}S_i = \{x : q_i(x) \le r_i\} conservatively contains every point where the renderer can return αi>1/255\alpha_i > 1/255. For Gaussian opacity oi>1/255o_i > 1/255, we use the outward-rounded radius ri=min⁡{radm,2ln⁡(255oi)}≤3r_i = \min\{r_{\mathrm{adm}},\sqrt{2\ln(255o_i)}\} \le3. If oi≤1/255o_i \le1/255, the support is empty, so the Gaussian may be rejected. 3DGUT’s renderer [31] uses a kernel cutoff of 0.0113, corresponding to radius −2ln⁡0.0113=2.994<3\sqrt{-2\ln0.0113}=2.994<3, so the cap radm=3r_{\mathrm{adm}}=3 contains its support. Complete admission omits the 0.2 near-plane cull of 3DGS [17] and 3DGUT.

For each Gaussian–view pair, we test every tile once. A tile is rejected only when one of three tests certifies that no valid ray can reach SiS_i; a failed or inconclusive test retains it. The Domain test rejects a tile with no pixel in Pv∩DvP_v \cap D_v, or when the angular cap of SiS_i seen from oo lies entirely beyond the calibrated maximum ray angle. For the Taylor test, suppose oo lies outside a ball B(μi,Ri)⊇SiB(\mu_i,R_i) \supseteq S_i. For the active view, let π\pi denote its projection and JπJ_\pi the projection Jacobian at μi\mu_i. We enclose π(Si)\pi(S_i) in a rectangle centered at π(μi)\pi(\mu_i) with half-extents

ha=ri∥(JπBi)a∥+ri∥(Jπ)a∥∥Bi∥η1−η+12HˉθRi2(∥μi−o∥−Ri)2+ϵa,(9)h_a = r_i\lVert(J_\pi B_i)_a\rVert+ r_i\lVert(J_\pi)_a\rVert\lVert B_i\rVert\frac{\eta}{1-\eta} + \frac{1}{2}\bar{H}_\theta\frac{R_i^2}{(\lVert\mu_i-o\rVert-R_i)^2} + \epsilon_a, \tag*{(9)}

where a∈{1,2}a \in\{1,2\} indexes the image coordinates. The bound Hˉθ\bar{H}_\theta is the outward-rounded interval bound on the Frobenius norm of the projection Hessian at unit range over the ray-angle band of SiS_i. The certified factor η=∥I−BiTBi∥<1\eta= \lVert I - B_i^\mathsf{T}B_i\rVert< 1 bounds the error in the renderer’s floating-point inverse AiA_i of BiB_i, while ϵa\epsilon_a encloses numerical errors in π\pi and JπJ_\pi. The test rejects a tile only if none of its pixel centers lies in this rectangle. For supports that cross the maximum ray angle, we apply the same bound in an angular chart. When the Taylor bound cannot be certified, the Cone test encloses all pixel-ray directions of the tile in a ball B(c,ϱ)B(c,\varrho). Choose r≥radmr \ge r_{\mathrm{adm}} to include a bound on the renderer’s floating-point error, and define p=Σi−1(μi−o)p = \Sigma_i^{-1}(\mu_i-o), κ=∥μi−o∥Σi−12−r2>0\kappa= \lVert\mu_i-o\rVert_{\Sigma_i^{-1}}^2-r^2>0, and K=ppT−κΣi−1K = pp^\mathsf{T}-\kappa\Sigma_i^{-1}. The tile is rejected if

cTp+ϱ∥p∥≤0orcTKc+2ϱ∥Kc∥+ϱ2∥K∥F<0.(10)c^\mathsf{T}p+\varrho\lVert p\rVert\le0 \quad\text{or}\quad c^\mathsf{T}Kc+2\varrho\lVert Kc\rVert+\varrho^2\lVert K\rVert_F<0. \tag*{(10)}

If the camera lies inside the enlarged support or cannot be certified outside it, every tile containing a valid pixel is retained. We retain 3DGUT’s unscented-transform parameters and renderer constants but replace its projected sigma-point extent with complete admission. Admitted Gaussians are ordered by radial distance using StopThePop’s per-pixel hierarchical sort [25].

Assumptions of Proposition 1. The camera is central, and every valid pixel satisfies π(o+t dv(u))=u\pi(o+t\,d_v(u))=u for all t>0t>0 in the chart. The angular cap encloses every direction from oo to SiS_i, the ball B(c,ϱ)B(c,\varrho) encloses every valid ray direction of the tile, and the rectangle of Eq. (9) contains π(Si)\pi(S_i) whenever o∉B(μi,Ri)o\notin B(\mu_i,R_i).

Proof of Proposition 1. Consider a valid pixel uu in the tile whose ray intersects SiS_i, and choose x=o+tdx=o+t d ∈Si\in S_i on that ray, with t≥0t\ge0. If t=0t=0, then o∈Sio\in S_i, so it is not provably outside the support, and every valid tile is kept. Otherwise t>0t>0. The Domain test cannot reject the tile because it contains uu, whose valid ray points to SiS_i and therefore lies inside the angular cap and calibrated range. The Taylor test cannot reject it because π(x)=u\pi(x)=u, so u∈π(Si)u\in\pi(S_i) and lies inside the rectangle. For the Cone test, write f(t)=qi(o+td)2=κ+r2−2t dTp+t2dTΣi−1df(t)=q_i(o+td)^2=\kappa+r^2-2t\,d^\mathsf{T}p+t^2d^\mathsf{T}\Sigma_i^{-1}d, where κ>0\kappa>0. For some t≥0t\ge0 to satisfy f(t)≤r2f(t)\le r^2, the quadratic requires dTp>0d^\mathsf{T}p>0 and (dTp)2≥κ dTΣi−1d(d^\mathsf{T}p)^2\ge\kappa\,d^\mathsf{T}\Sigma_i^{-1}d, or equivalently dTKd≥0d^\mathsf{T}Kd\ge0. Because d∈B(c,ϱ)d\in B(c,\varrho), we have dTp≤cTp+ϱ∥p∥d^\mathsf{T}p\le c^\mathsf{T}p+\varrho\lVert p\rVert and dTKd≤cTKc+2ϱ∥Kc∥+ϱ2∥K∥Fd^\mathsf{T}Kd\le c^\mathsf{T}Kc+2\varrho\lVert Kc\rVert+\varrho^2\lVert K\rVert_F. Either rejection condition in Eq. (10) would therefore prevent dd from reaching qi≤rq_i \le r. But x∈Six \in S_i gives qi(x)≤ri≤radm≤rq_i(x) \le r_i \le r_{\mathrm{adm}} \le r, so the Cone test cannot reject the tile.

Calibrated Mip Filtering

3D filter. A training view observes Gaussian ii when πv(μi)∈Dv\pi_v(\mu_i) \in D_v. We detach the mean and Jacobians when computing ϵi\epsilon_i, then freeze it with the scene when geometry optimization ends. Following Mip-Splatting [36], a Gaussian unseen by every training view receives the largest ϵi\epsilon_i among the observed Gaussians. Proposition 3 shows that Mip-Splatting’s 3D volume-ratio scaling does not generally preserve ray-maximum image integrals, so we leave opacity unchanged. Mesh extraction uses the filtered covariance and optimized opacity oio_i, while the per-view factor mi,vm_{i,v} is applied only during rendering.

Pixel footprint. We detach the mean and Jacobian when computing the lift of Eq. (5), so mi,vm_{i,v} creates no position gradient along the viewing ray. For pinhole cameras, Ji,vJ_{i,v} is the analytic perspective Jacobian, and we skip the footprint below a view-space depth of 0.2. Other cameras require a full-row-rank Jacobian at the Gaussian mean within the valid camera chart. If none exists, we set Ci,v=ΣiC_{i,v} = \Sigma_i and mi,v=1m_{i,v}=1; admission still uses the covariance rendered for that view. We compute Ji,v+J_{i,v}^{+} in double precision using a pivoted QR factorization of Ji,vJ_{i,v} because normal equations lose accuracy at grazing angles.

Proof of Proposition 2. Let the columns of U∈R3×2U \in\mathbb{R}^{3\times2} form an orthonormal basis for the row space of JJ, and let nn be a unit null vector. In the basis [U n][U\ n], write XX in blocks Y11Y_{11}, y12y_{12}, and y22y_{22}. Because Jn=0Jn = 0, the constraint becomes (JU)Y11(JU)T=γI2(JU)Y_{11}(JU)^{\mathsf{T}} = \gamma I_2 and uniquely fixes Y11=γ(JU)−1(JU)−TY_{11} = \gamma(JU)^{-1}(JU)^{-\mathsf{T}}. Positive semidefiniteness gives y22≥0y_{22} \ge0. Since tr⁡X=tr⁡Y11+y22\operatorname{tr} X = \operatorname{tr}Y_{11} + y_{22}, the minimum requires y22=0y_{22}=0, which in turn forces y12=0y_{12}=0. Thus the unique minimizer is UY11UTUY_{11}U^{\mathsf{T}}. Using J+=U(JU)−1J^{+} = U(JU)^{-1}, this equals γJ+J+T\gamma J^{+}J^{+\mathsf{T}} and has nTXn=0n^{\mathsf{T}}Xn=0.

Opacity normalization. Minimizing Mahalanobis distance along a ray, as in Eq. (1), yields the same transverse profile as integrating the Gaussian along that ray, up to the factor 2πs\sqrt{2\pi}s defined below. Under a local projection, this response is a 2D Gaussian with covariance Ji,vCi,vJi,vTJ_{i,v}C_{i,v}J_{i,v}^{\mathsf{T}}. When the αmax⁡\alpha_{\max} cap is inactive, its integral over the tangent image plane is proportional to oidet⁡(Ji,vCi,vJi,vT)1/2o_i \det(J_{i,v}C_{i,v}J_{i,v}^{\mathsf{T}})^{1/2}. Preserving this integral as Σi\Sigma_i grows to Ci,vC_{i,v} gives mi,vm_{i,v} in Eq. (5). The 3D volume ratio (det⁡Σi/det⁡Ci,v)1/2(\det\Sigma_i/\det C_{i,v})^{1/2} instead preserves the 3D integral.

Proof of Proposition 3. Choose an orthonormal frame with one axis along dd and two transverse axes, and write Σ\Sigma in blocks Σ⊥⊥\Sigma_{\perp\perp}, Σ∥∥=Σ∥⊥T\Sigma_{\parallel\parallel} = \Sigma_{\parallel\perp}^{\mathsf{T}}, and Σ∥∥\Sigma_{\parallel\parallel}. The determinant factorizes as det⁡Σ=det⁡(Σ⊥⊥)s(Σ)\det\Sigma= \det(\Sigma_{\perp\perp})s(\Sigma), where the Schur complement is positive definite and conditioned on transverse position. Because Jd=0Jd = 0, the determinant det⁡(JΣJT)\det(J\Sigma J^{\mathsf{T}}) is proportional to det⁡Σ⊥⊥\det\Sigma_{\perp\perp}, with the same proportionality for CC. Hence mi,v2=det⁡Σ⊥⊥/det⁡C⊥⊥m_{i,v}^{2} = \det\Sigma_{\perp\perp}/\det C_{\perp\perp}, which gives the identity. The lift of Proposition 2 changes only the transverse block. Writing its positive-definite transverse component as Δ\Delta, we obtain

s(C)−s(Σ)=Σ∥∥[Σ⊥⊥−1−(Σ⊥⊥+Δ)−1]Σ∥∥≥0.s(C)-s(\Sigma)=\Sigma_{\parallel\parallel}\left[\Sigma_{\perp\perp}^{-1}-(\Sigma_{\perp\perp}+\Delta)^{-1}\right]\Sigma_{\parallel\parallel}\ge0.

Because Δ≻0\Delta\succ0, the bracket is positive definite, so equality holds exactly when Σ∥∥=0\Sigma_{\parallel\parallel}=0, which is equivalent to dd being an eigenvector of Σi\Sigma_i.

Optimization

Objective. Following 3DGS [17], GOFish uses the photometric loss 0.8L1+0.2LD−SSIM0.8\mathcal{L}_1+0.2\mathcal{L}_{\mathrm{D-SSIM}}. The Geometry Objective of § 6.5 adds SOF’s terms [26] with their original weights and warm-ups after adapting camera-dependent quantities to the calibrated ray domain,

L=0.8L1+0.2LD−SSIM+100Ldist+0.05Lnormal+0.01Lsmooth+0.004Lopa+0.1Lext.(11)\mathcal{L}=0.8\mathcal{L}_1+0.2\mathcal{L}_{\mathrm{D-SSIM}}+100\mathcal{L}_{\mathrm{dist}}+0.05\mathcal{L}_{\mathrm{normal}} +0.01\mathcal{L}_{\mathrm{smooth}}+0.004\mathcal{L}_{\mathrm{opa}}+0.1\mathcal{L}_{\mathrm{ext}}. \tag*{(11)}

Following GOF [37], the distortion loss fixes compositing weights and differentiates hit depths only through Gaussian positions and shapes; differentiating the weights as in SOF gives nearly identical results on FLORD. We use no pose optimization or transient modeling. On ScanNet++, we add CoMe’s confidence-weighted photometric loss and normal-variance regularizer [27] with their released weights from iteration 500. We omit its confidence-steered densification and color-variance loss, which would compete with PPISP. Appendix C.1 reports these losses on every dataset.

Opacity and depth–normal terms. SOF’s opacity loss Lopa\mathcal{L}_{\mathrm{opa}} drives accumulated ray opacity toward 0.5 at the surface crossing. As in its released code, we solve t0.5t_{0.5} in closed form on the rising side of the Gaussian that takes opacity through 0.5 and hold it fixed during backpropagation. A second pass evaluates the Gaussian scene without the pixel footprint, stopping at t0.5t_{0.5} or traversing the full ray when no crossing exists. The depth-normal loss Lnormal\mathcal{L}_{\mathrm{normal}} compares rendered normals with normals computed from the detached depth at t0.5t_{0.5} wherever opacity reaches 0.5. Unlike SOF, which passes a straight-through gradient through this depth, GOFish detaches it. We retain SOF’s edge-aware normal smoothness loss Lsmooth\mathcal{L}_{\mathrm{smooth}} unchanged.

Distortion and extent terms. These terms require two camera-domain adaptations of SOF’s objective. Let ti∗t_i^{*} be a hit’s peak parameter along a unit ray whose direction has camera-frame optical-axis component dzd_z, and define

ui={ti∗dz,pinhole,ti∗,fisheye,ξi=1max⁡(ui,0.2).(12)u_i = \begin{cases} t_i^{*}d_z, & \text{pinhole},\\ t_i^{*}, & \text{fisheye}, \end{cases} \qquad \xi_i = \frac{1}{\max(u_i,0.2)}. \tag*{(12)}

For compositing weights wi=αiTiw_i=\alpha_iT_i, where TiT_i is the incoming transmittance, define

S0=∑iwi,S1=∑iwiξi,S2=∑iwiξi2,(13)S_0=\sum_i w_i, \qquad S_1=\sum_i w_i\xi_i, \qquad S_2=\sum_i w_i\xi_i^2, \tag*{(13)}
Ldist=cndc2(S0S2−S12),\mathcal{L}_{\mathrm{dist}}=c_{\mathrm{ndc}}^2(S_0S_2-S_1^2),

where cndc=tftn/(tf−tn)c_{\mathrm{ndc}}=t_ft_n/(t_f-t_n) is SOF’s normalized-device-coordinate scale for far and near planes tf=100t_f=100 and tn=0.2t_n=0.2. The clamp applies only to this loss coordinate and does not affect rendering. Pinhole hits above the near plane match SOF’s coordinate exactly. The extent term accumulates each Gaussian’s alpha-weighted along-ray standard deviation (d⊤Σid)1/2(\mathbf{d}^{\top}\Sigma_i\mathbf{d})^{1/2} in world units. This removes the near-camera bias of normalized device coordinates and penalizes oversized Gaussians viewed obliquely. As in SOF’s released code, the compositing alphas are detached in this term, and we keep SOF’s weight 0.1 for it on every dataset.

Densification. For a full-row-rank Jacobian, we recover the image gradient as gu,i,v=Ji,v+⊤gμ,i,v\mathbf{g}_{u,i,v}=J_{i,v}^{+\top}\mathbf{g}_{\mu,i,v} and scale it to normalized device coordinates as in SOF. For AbsGS [33], we accumulate the absolute gradient components alongside SOF’s L2L_2 norm. We densify a Gaussian when its signed image-space gradient exceeds 2×10−42\times10^{-4} or its absolute gradient exceeds a quantile chosen, following GOF, to select the same fraction of Gaussians. Densification and pruning run every 100 iterations from 500 to 15k, and opacities are reset to 0.01 every 3k iterations as in 3DGS.

Pruning. We retain SOF’s opacity threshold of 0.05 and prune Gaussians larger than 0.1 of the scene extent. On FIORD [8] and Tanks & Temples [18], we use 3DGUT’s robust extent, 1.1 times the median camera distance from the coordinatewise-median camera center, so one distant capture cannot inflate it. On ScanNet++ [34], we use SOF’s extent, 1.1 times the maximum camera distance from the camera centroid. The median rule fails on ScanNet++ 0a5c, whose compact 2 m camera path gives a 4.6 cm threshold that deletes 40% of the Gaussians and leaves near-camera floaters. Conversely, the maximum rule retains large Gaussians that create off-surface mesh on FIORD and Tanks & Temples. We omit 3DGS’s 20-pixel screen-size prune, inherited by SOF, GOF, and CoMe, because densification resets the screen radii before it can run.

Appearance. The Gaussian scene and PPISP train jointly for 30k iterations, after which only PPISP’s novel-view controller trains for 5k iterations. We extract the mesh from the 30k scene. Beyond adding the wide-angle attenuation WW of Eq. (7), we modify PPISP only so its pooling and gating ignore pixels outside the calibrated domain. For noncircular fisheye images such as ScanNet++, the image corner defines the maximum image radius. The radial profile ν\nu has 50 knots over the outer half of this radius, and the flat field aa is a 50×5050\times50 grid. Both are initialized to unit attenuation and use a smoothness weight of 10−410^{-4}.

Mesh Extraction

Following GOF, we place nine sites per Gaussian and build a fresh Delaunay tetrahedralization over all sites. We evaluate OO of Eq. (8) at each site and extract the level set O=0.5O=0.5 with Marching Tetrahedra. Let RiR_i and Σi\Sigma_i denote Gaussian ii’s rotation and scale before filtering. The isotropic 3D filter preserves its rotation and gives Σi=RiSˉiSˉi⊤Ri⊤\Sigma_i=R_i\bar{S}_i\bar{S}_i^{\top}R_i^{\top}, where Sˉi2=Si2+ϵi2I3\bar{S}_i^2=S_i^2+\epsilon_i^2I_3 is evaluated elementwise on the diagonal. The sites are μi\mu_i and the eight points μi+riRiSˉiz\mu_i+r_iR_i\bar{S}_i\mathbf{z} for z∈{±1}3\mathbf{z}\in\{\pm1\}^3. Here rir_i is defined in Appendix A.2 and capped at 3, matching GOF’s box for sufficiently opaque Gaussians. Sites are placed only for Gaussians with oi>1/255o_i>1/255.

Field evaluation. Mesh extraction uses a separate practical admission rule, the union of 3DGUT’s projected sigma-point extent and a tile box enclosing the projected angular cap of the support. As in GOF, view validity in Eq. (8) checks only the calibrated camera domain, not occlusion. We refine each Marching Tetrahedra crossing with eight bisection steps on the same field, where unobserved midpoints remain opaque as in GOF and SOF.

Filters. Before Marching Tetrahedra, we apply the cell filter described by GOF [37] and remove any tetrahedron with an edge longer than the sum of its endpoint Gaussians’ support extents. Neither the GOF nor SOF release implements this filter. We then apply SOF’s released filter, which retains a crossing vertex when its generating edge is no longer than the sum of the endpoint scales, then retains a face only if all three vertices remain. The cell filter does not determine the comparison, since without it GOFish still reaches an F-score of 0.244 on FIORD against 0.239 for PGSR [2].

Real-Time Renderer

GOFish’s reported FPS uses the real-time renderer described here, while training and every tabulated rendering metric use the complete-admission renderer of § 6.1. The real-time renderer uses the same trained Gaussians, filtering, and PPISP model, recomputing PPISP each frame while caching camera-fixed terms such as WW.

Gaussian admission starts from 3DGUT’s projected sigma-point extent, with out-of-domain sigma points mirrored from valid counterparts. If the extent remains undefined and the Gaussian subtends more than 5∘5^\circ, a conservative cone test admits it. The renderer sorts admitted Gaussians per tile and composites them through a 12-entry per-pixel buffer. Tiles containing more than 2048 Gaussians are divided among four thread blocks.

With complete admission and per-pixel sorting enabled, it reproduces the training renderer byte for byte. Relative to the complete-admission renderer, it lowers Full PSNR by 0.03 dB on FIORD, 0.14 dB on FullCircle, and 0.002 dB on ScanNet++ without changing any first-place Full ranking.

Additional Experimental Details

FIORD Data

FIORD [8] provides full-FoV fisheye images with metric ground truth. Figure 8 shows the excluded SfM trajectories, Corridor_Out’s re-registration, and field-of-view calibration. We render seven of its ten scenes and evaluate geometry on six after method-independent registration and calibration checks.

FIORD capture and calibration diagnostics

Figure 8. FIORD capture and calibration diagnostics [8]. (Top) SfM reconstructions are overlaid on LiDAR (gray). Points within 10 cm of LiDAR are green and others orange. Night_Out has no LiDAR. Cameras are black; red marks those more isolated than any retained camera, including one off-panel (arrow). (a) Full-FoV 95th-percentile training luma (left) and a rim crop (right) for Building_In camera 2. Curves mark total field angles of 167° (blue) and 200° (orange), while magenta marks the dark band. (b) Held-out ground truth (left) and GOFish rendering (right) at the rim, with the calibrated edge in white.

Excluded scenes. The authors’ SfM reconstructions of Night_Out, Building_Out, and Upstairs_In do not form coherent camera trajectories. Every retained camera is within 0.29 SfM units of another; the most isolated cameras in these scenes are 1.21, 0.96, and 100.6 units from the rest, respectively. In Upstairs_In, the outlier is supported by only 33 of the scene’s 22,247 keypoints and lies 234 m from the remaining trajectory. We therefore exclude all three; Night_Out also lacks LiDAR ground truth.

Registration. Corridor_Out’s images and poses are internally consistent and retained for rendering, but its SfM–LiDAR registration is invalid. Global re-registration with Fast Point Feature Histograms (FPFH), Random Sample Consensus (RANSAC), and robust Iterative Closest Point (ICP) reduces the camera-up tilt from 14.6° to 1.7° but leaves an inlier root-mean-square error of 0.114 m, above our 0.10 m limit. We therefore exclude Corridor_Out from the geometry mean and report it only for rendering. A method-independent ICP alignment between Hall_In’s SfM points and LiDAR consistently corrects its 3.4° tilt, and we apply it to all methods.

Field-of-view calibration. The published FIORD intrinsics are extrapolated at the image-circle boundary. On six cameras, the calibration maps this boundary to total FoVs of 198° to 244° for a nominal 200° lens. On the other eight, it loses monotonicity before the boundary, which ends the valid domain between 186° and 229°. Six of the 14 cameras contain a dark band at calibrated total field angles of at least 167°, covering 3.1–5.7% of the valid image domain. Although the calibration assigns scene rays to this band, its 95th-percentile training luma remains below 0.10 because it receives no scene light in any training frame. A static scene cannot reproduce this camera-fixed darkness, so the per-camera attenuation WW of § 6.4 absorbs it without entering the scene. We retain the full calibrated FoV, including this band, when scoring every method. Excluding it lowers Full PSNR by 0.07 dB for the Geometry Objective variant and 0.04 dB for DirectFisheye-GS [32].

Evaluation Protocol

Datasets. Camera-native methods train on 816×816 FIORD images and 720×720 FullCircle images per lens [5], both downsampled by 4×. ScanNet++ [34] uses 876×584 DSLR fisheye images downsampled by 2×, while Tanks & Temples [18] uses images near 980×545 from GOF’s preprocessing [37]. Rectified methods use matching 800×800 images for FIORD and FullCircle and the official undistorted 876×584 images for ScanNet++. Every eighth image is held out for testing, except FullCircle, which uses separately captured tripod frames. ScanNet++ uses the six scenes of 3DGUT and Fisheye-GS [22, 31], and Tanks & Temples uses GOF’s six scenes.

Rendering. We use no test-time color, exposure, or pose fitting and leave learned image processors unchanged. Full is the calibrated FoV of Appendix § 12.§ 1, excluding only the tripod pixels of FullCircle’s test mask. Common is the rectified training raster; camera-native predictions are resampled onto it with the fixed rectification operator.

Camera-native methods use their released renderers. For rectified methods, we port each rasterizer to the calibrated lens while preserving its activations, filters, sorting, alpha rules, appearance handling, and background, with no GOFish component. On rectified test views, every port matches its released renderer within 0.01 dB PSNR, 0.0002 SSIM, and 0.0004 LPIPS [38] on every scene. SOF and CoMe [26, 27] match byte for byte. For each FullCircle scene, we use whichever of two GOF ports, differing only in numeric precision, passes this check.

To avoid renderer conversion, the GOFish @ FoV columns use each baseline’s released renderer on its own domain and score GOFish on the same pixels. That domain contains pixels whose ray angle is at most half the renderer’s listed FoV. SSIM uses an 11×11 Gaussian window and LPIPS uses VGG, both restricted to each scored region; on the GOFish @ FoV domains, LPIPS is averaged over 64×64 tiles. On Tanks & Temples, each method uses its released pinhole renderer; when dimensions differ, we resize only the prediction with a fixed operator.

Geometry. Every baseline retains its released mesh post-processing; GOFish uses Appendix § 12.§ 7. On FIORD, observed ground truth comprises LiDAR points that hidden-point removal [16] finds visible within the calibrated domain from at least two training cameras. Recall counts a point as recovered when its nearest prediction is within 10 cm. As on Tanks & Temples and ScanNet++, precision is evaluated only in regions supported by the ground truth. Predicted samples must lie inside the LiDAR’s axis-aligned bounding volume and be observable by the same test front from at least one training camera. Both visibility selections depend only on the ground truth and training cameras. On ScanNet++, every method follows CoMe’s protocol of scaled ICP per mesh, cropping to the ground-truth volume, 2.5 cm voxel downsampling, and a 5 cm threshold. On Tanks & Temples, we use the official crop, alignment, thresholds, and evaluator.

Runtime. All times are wall-clock measurements on one NVIDIA A10G GPU. Training includes the full schedule and the novel-view controller phase of PPISP [4]; meshing includes extraction and post-processing of the 30k scene. Rendering speed is measured at each method’s evaluation resolution over 200 synchronized held-out frames after 20 warm-ups, timing only the trained-scene renderer; GOFish uses Appendix § 12.§ 8. GOFish’s training times use an optimized build with a bit-identical forward pass and gradients that agree up to rounding; metrics use standard runs.

Component analysis. In Table 3, F averages the six geometry scenes and Train excludes meshing. The + Full-FoV Rasterization row keeps 3DGUT’s densification, and from + Calibrated Mip Filtering on, rows use the calibrated densification gradient and the schedule of Appendix § 12.§ 6. Every 3DGUT and 3DGRT [24] row uses the same opacity-field readout, and empty extractions score 0. The + Full-FoV Rasterization row includes only exact speedups and reports FPS from its own configuration without the Mip filter.

Baseline Adaptations

On FIORD, SOF [26] uses its released large-scene recipe at the rectified resolution; on Tanks & Temples, PGSR’s released preprocessing [2] supplies scene bounds absent from GOF’s inputs. MiLo [7] exceeds the 24 GB memory of the NVIDIA A10G GPU on FIORD Corridor_Out and Road_Out, so we train those scenes with the same released configuration on a 40 GB NVIDIA A100 GPU.

Fisheye-GS [22] supports only an equidistant fisheye model. On FIORD and ScanNet++, we reproject training images to its equidistant camera and map held-out predictions back to the native raster for scoring. Pixels outside the source image use zeros on FIORD and reflected borders on ScanNet++. On Tanks & Temples, we make its training script honor pinhole mode. We also restore the forward-pass projection clamp at 1.3tan⁡(FoV/2)1.3\tan(\mathrm{FoV}/2) assumed by its backward pass, as in 3DGS [17]. Without the clamp, Gaussians far outside the frustum render without the intended bound while receiving clamped gradients, producing floaters in held-out views.

Method\multicolumn{9}{c}{ScanNet++ (3DGUT set) [31, 34]}\multicolumn{9}{c}{ScanNet++ (CoMe set) [27, 34]}
\multicolumn{3}{c}{Full (127°)}\multicolumn{3}{c}{Common}\multicolumn{3}{c}{Geometry}\multicolumn{3}{c}{Full (127°)}\multicolumn{3}{c}{Common}\multicolumn{3}{c}{Geometry}
MethodPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PP ↑RR ↑FF ↑PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PP ↑RR ↑FF ↑
GOF [37]27.540.9040.21230.010.9200.2080.5790.6290.60126.100.8740.23427.340.8890.2330.5780.6930.630
PGSR [2]27.390.8980.24129.660.9160.2290.6980.6770.68725.690.8700.25826.940.8870.2490.6940.7280.710
SOF [26]27.100.8980.21828.830.9140.2140.6200.6670.64126.290.8760.22127.200.8890.2250.6030.7240.657
MiLo [7]27.610.9030.23429.750.9160.2310.6510.7010.67526.350.8770.24727.500.8910.2470.6520.7260.687
CoMe [27]28.110.9090.19029.920.9210.1930.6520.7300.68726.650.8800.20727.550.8930.2110.6310.7680.692
GOFish (Ours)30.170.9210.17230.720.9250.1850.6370.7950.70627.610.8880.19528.240.8990.2080.6040.8280.698
+ Geometry Objective30.120.9210.17130.680.9250.1850.6300.8010.70427.610.8890.19328.200.8990.2070.5920.8460.696

Table 4. Quantitative comparison on both ScanNet++ scene sets [34]. Left uses the six scenes of 3DGUT and Fisheye-GS [22, 31]; right uses CoMe’s six [27]. Full is the native 127∘127^\circ image and Common the official undistortion. GOFish is mapped to Common with the official map and a 2×2\times box downsample; geometry uses CoMe’s 5 cm protocol. The best, second best, and third best entries are highlighted.

Self-Calibrating GS [3] initializes its lens network from a fisheye calibration file. For the pinhole Tanks & Temples images, we supply a zero-distortion pinhole calibration, so the lens starts at identity rather than its equidistant fallback. On ScanNet++, its 90∘90^\circ cube-face guard rescales only the vertical focal length, although the same intrinsics define the fisheye-to-cube mapping. We retain the calibrated focal length for that mapping and keep the face intrinsics consistent in the evaluation cameras. We also use the dataset’s Kannala–Brandt calibration instead of its equidistant default for the base lens mapping. Training does not save the cubemap network, so the method’s own initialization refits it before the released held-out render pass used for timing.

3DGUT and 3DGRT [24, 31] use a released loader that sets FullCircle’s field-of-view limit from the image corner. The corner lies beyond the calibration’s monotone range, producing folded rays and collapsing 3DGUT training on room3 to 11.7 dB full PSNR. We cap the limit at the end of the monotone range and assign no ray beyond it, matching FullCircle’s treatment of border pixels.

3DGEER [14] uses KB mode on FIORD and FullCircle and BEAP training with KB rendering on ScanNet++, following its released configurations at our splits and resolutions. Because its fisheye models assume forward-facing rays, we use the release’s mask to exclude rays at least 89∘89^\circ from the optical axis, limiting its field of view to 178∘178^\circ. These rays receive no loss and are scored as unrendered black pixels, as for every baseline. We supply one ray map per FIORD camera, computed with a converged inverse of the KB model, because the release’s three-step Newton inverse does not converge on FIORD’s lenses. We also cull Gaussians whose KB pixel bound is inverted for a view, a case the release does not catch and that crashes training on two outdoor scenes. On FullCircle lounge, 3DGEER exceeds the memory of our 24 GB and 32 GB GPUs at its released densification threshold, so we double it to 0.0000 and leave all other settings unchanged.

MethodPSNR↑SSIM↑LPIPS↓PP ↑RR ↑FF ↑
\multicolumn{1}{c}{}\multicolumn{3}{c}{Rendering}\multicolumn{3}{c}{Geometry}
Fisheye-GS* [22]24.710.8380.236
3DGUT [31]25.110.8520.221
3DGUT + PPISP [31, 4]26.510.8680.199
3DGEER* [14]24.420.8370.235
Self-Calibrating GS* [3]22.660.7590.332
DirectFisheye-GS [32]25.370.8460.227
3DGRT [24]23.950.8420.232
3DGRT + PPISP [24, 4]26.420.8630.200
GOF [37]22.850.8330.2310.5110.4170.451
PGSR [2]23.630.8400.2120.4760.5850.514
SOF [26]22.960.8140.2480.5200.4220.454
MiLo [7]22.960.8130.2810.5220.3960.445
CoMe [27]22.450.8140.2450.5740.4950.521
GOFish (Ours)26.770.8720.1870.3890.4220.399
+ Geometry Objective26.520.8680.1930.4060.4350.414

Table 5. Quantitative comparison on Tanks & Temples [18]. The best, second best, and third best entries are highlighted. * Appendix B.3 details baseline adaptations.

Additional Results

Table 4 compares ScanNet++ [34] on the six scenes used by 3DGUT and Fisheye-GS [22, 31] and the six used by CoMe [27]. Across its base and Geometry Objective variants, GOFish leads rendering and recall on both sets. Base GOFish also leads F-score on the 3DGUT set and ranks second on CoMe’s set, 0.698 against 0.710 for PGSR [2]. On the pinhole Tanks & Temples benchmark [18] in Table 5, GOFish renders best, while dedicated surface methods retain higher F-score. Figure 15 shows every scene; on Courthouse, floaters cover CoMe’s rendering while GOFish recovers the facade.

Qualitative rendering and mesh results for Barn, Caterpillar, Courthouse, Ignatius, Meetingroom, and Truck

Figure 15. Qualitative rendering and mesh results on Tanks & Temples [18]. Metrics are per-scene.

Ablations and Robustness

Cross-dataset losses. Across separate full training runs, CoMe’s confidence-weighted photometric loss and normalvariance regularizer raise F-score by 0.011, 0.028, and 0.013 on FIORD [8], ScanNet++, and Tanks & Temples. Full PSNR changes by −0.13, +0.06, and −0.11 dB, respectively. Following CoMe’s ScanNet++ protocol, GOFish uses these losses only there. The cross-dataset runs show that both metrics improve on ScanNet++, whereas modest F-score gains on FIORD and Tanks & Temples come with lower Full PSNR.

Run-to-run variation. Each of three repetitions covers every evaluated FIORD scene for both GOFish and PGSR. Across the six geometry scenes, base GOFish averages an F-score of 0.245 ± 0.001 versus 0.239 ± 0.002 for PGSR, and every GOFish run exceeds every PGSR run. Its seven-scene mean Full PSNR varies by only ±0.02 dB. The F-score gap therefore exceeds run-to-run variation.

Densification split. Table 3 adds calibrated densification and filtering together. We separate them on four of FIORD’s six geometry scenes [8]: Kitchen_In, Road_Out, MeetingRoom_In, and Bridge_Out. The Building_In runs used an earlier training configuration, while unfiltered Hall_In training ran out of GPU memory. With only the image-space statistic, only SOF’s schedule [26], or only the filter, the mean F-score is 0.146, 0.214, and 0.140, against 0.301 for the full row on these scenes. On Kitchen_In and MeetingRoom_In, where both variants produce meshes, calibrated densification without the filter scores 0.573 and 0.503, against 0.560 and 0.505 with it.

With and without the filter, Full PSNR differs by at most 0.1 dB on these four scenes and on Corridor_Out, which has no geometry ground truth. The filter reduces the Gaussian count by 2.2–5.4×, from 1.05–2.94 M to 0.42–0.69 M, and raises real-time rendering speed on Kitchen_In and MeetingRoom_In from 65 to 86–90 FPS on an NVIDIA A10G GPU. Without the filter, Bridge_Out mesh extraction ran out of GPU memory, while Road_Out mesh extraction did not finish. Thus calibrated densification drives surface recovery, while filtering controls scene size and memory.

Geometry by region. We assign each FIORD ground-truth point and predicted sample the smallest ray angle from which any training camera observes it. We separately identify samples observed by any camera within 90° rectification. This rectified region contains 94% of the observed ground truth. GOFish leads every surface method within the shared 90° FoV, 0.254 against 0.243 for PGSR, and for rays beyond 90° from the optical axis, 0.112 against 0.102 for CoMe. Its overall advantage therefore does not come only from evaluating the wider FoV.

Efficiency

Table 6 separates training, meshing, and rendering cost from representation size. On FIORD, GOFish completes training and mesh extraction in 153 minutes, faster end to end than GOF’s 208 minutes [37]. Its real-time renderer reaches 79 FPS, close to the 84 FPS of rendering-only 3DGUT + PPISP [4, 31] and more than three times the 22 FPS of 3DGRT + PPISP [24]. The same renderer reaches 99 FPS on ScanNet++ and 82 FPS on Tanks & Temples. Despite covering the full FoV, GOFish averages 0.60 million Gaussians on FIORD, compared with 1.03 million for 3DGUT + PPISP. The optional Geometry Objective adds little end-to-end cost and does not reduce rendering speed.

MethodFIORD [8]ScanNet++ [34]Tanks & Temples [18]
Train↓Mesh↓Total↓FPS↑GaussiansTrianglesTrain↓Mesh↓Total↓FPS↑GaussiansTrianglesTrain↓Mesh↓Total↓FPS↑GaussiansTriangles
3DGS [17]23m23m1240.95M16m16m1680.97M22m22m1071.31M
Fisheye-GS⋆ [22]42m42m1350.60M119m119m74.00.96M21m21m1151.49M
3DGUT [31]28m28m92.11.03M10m10m2380.37M24m24m1051.00M
3DGUT + PPISP [31, 4]29m29m84.31.03M11m11m2090.35M25m25m96.51.00M
3DGEER⋆ [14]51m51m46.71.90M67m67m1121.41M71m71m97.24.27M
Self-Calibrating GS⋆ [3]123m123m7.00.51M98m98m6.80.34M84m84m30.72.03M
DirectFisheye-GS [32]28m28m2080.46M56m56m1370.66M29m29m1910.97M
3DGRT [24]104m104m23.11.03M27m27m95.70.53M63m63m39.81.00M
3DGRT + PPISP [24, 4]108m108m22.31.03M28m28m93.50.50M67m67m37.21.00M
GOF [37]104m104m208m11.00.73M9.9M76m51m127m18.20.98M16.7M93m46m139m11.71.34M28.3M
PGSR [2]49m309m358m1030.53M20.6M45m1m46m1180.60M4.9M69m4m73m43.61.44M22.9M
SOF [26]37m10m47m53.40.85M8.5M34m9m43m52.50.99M16.8M41m10m51m42.61.35M28.2M
MiLo [7]62m⋆6m⋆68m⋆3120.16M4.8M77m7m85m3140.21M6.1M94m10m104m1940.23M7.0M
CoMe [27]41m9m50m40.20.69M12.0M43m11m54m34.81.20M21.0M55m14m69m29.11.79M29.0M
GOFish (Ours)107m46m153m79.30.60M19.3M129m29m158m98.80.97M26.7M99m22m121m81.51.32M41.4M
+ Geometry Objective116m45m161m86.80.60M18.6M135m28m163m1070.97M26.5M103m21m124m84.01.32M39.6M

Table 6. Efficiency and representation size. Dataset averages follow Appendix § 13.§ 2. Training, meshing, and FPS are measured on an NVIDIA A10G GPU. GOFish FPS uses the real-time renderer detailed in Appendix § 12.§ 8. ⋆ denotes an NVIDIA A100 GPU run detailed in Appendix § 13.3.

Surface Failure Cases

Figure 9 shows challenges shared across Gaussian surface reconstruction methods [2, 7, 26, 37]. Reflective surfaces and thin structures challenge the minimum-over-views field in panels (a) and (b), while sparse observations produce holes and overfill in panels (c) and (d).

Representative surface-reconstruction failure cases

Figure 9. Representative surface-reconstruction failure cases. Panels (a)–(c) compare ground truth (left) with GOFish meshes (right). Panel (d) shows GOFish’s rendering (left) and mesh (right).

Per-Scene Metrics

Tables 7–10 provide per-scene results for the aggregates in Tables 1, 2, 4, and 5. They also include FullCircle surface-baseline results and Tanks & Temples rendering and geometry that the body tables do not report. Rendering domains follow Appendix § 13.§ 2: Full is the calibrated FoV, Common the shared rectification, and GOFish @ FoV compares GOFish on the pixels rendered by each baseline. Blank cells denote unreported domains, and asterisks mark baseline adaptations detailed in Appendix § 13.3.

Additional Qualitative Results

Figures 10–15 show one held-out view per scene, emphasizing scenes not shown in the body. Every dataset has a rendering comparison; FIORD [8], ScanNet++ [34], and Tanks & Temples [18] also have surface comparisons, while FullCircle [5] lacks geometry ground truth. Badges report per-scene metrics. Meshes are rendered from the held-out camera and shaded as in Figure 5; black marks missing surfaces. GOFish uses the Geometry Objective for meshing.

DatasetAbbreviationScene
FIORD [8]Kit.Kitchen_In
Cor.Corridor_Out
RoadRoad_Out
Bld.Building_In
Meet.MeetingRoom_In
Bri.Bridge_Out
HallHall_In
FullCircle [5]Darkdark
F1flat1
F2flat2
Lablab
Lou.lounge
Per.persons
R1room1
R2room2
R3room3
ScanNet++ [34]0a5c0a5c013435
8d568d563fc2cc
bb87bb87c292ad
d415d415cc449b
e8eae8ea9b4da8
fe17fe1733741f
Tanks & Temples [18]BarnBarn
Cat.Caterpillar
Court.Courthouse
Ign.Ignatius
Meet.Meetingroom
TruckTruck

Table 11. Scene abbreviations for Tables 7–10.

Table 7. Per-scene rendering results on FIORD [8] and FullCircle [5] for fisheye baselines. The best, second best, and third best entries are highlighted. GOFish’s gain at each baseline’s FoV is highlighted. Protocols follow Appendix § 13.§ 2.

MetricMethodFoVFIORD Full (200°): Kit.Cor.RoadBld.Meet.Bri.HallFIORD GOFish @ FoV: Kit.Cor.RoadBld.Meet.Bri.HallFullCircle Full (205°): DarkF1F2LabLou.Per.R1R2R3FullCircle GOFish @ FoV: DarkF1F2LabLou.Per.R1R2R3
PSNR↑Fisheye-GS* [22]145°18.7216.2217.1918.1816.2817.4720.42+5.66+2.93+2.59+6.56+9.65+3.43+4.6313.4413.8714.5314.3812.8013.2514.9714.5914.28+15.05+13.98+14.05+14.29+8.24+15.82+15.51+15.64+13.88
PSNR↑3DGUT [31]160°20.2017.6219.4919.0319.6224.2518.84+3.99+4.01+2.48+6.36+5.27−1.51+7.0422.1721.0620.4420.3220.1422.5522.3023.2122.06+1.40+2.69+1.15+1.71−1.46+2.63+2.13+2.68+1.63
PSNR↑3DGUT + PPISP [31, 4]160°20.3418.7819.7019.6819.4824.3818.93+4.36+3.26+2.14+6.51+6.12−1.17+7.0922.0320.9920.5820.3120.0022.0622.0923.0721.93+1.74+2.53+1.44+1.82−1.42+3.13+2.84+3.00+1.82
PSNR↑3DGEER* [14]178°14.1910.3911.5212.2913.5010.0617.55−0.39+4.80+3.55+1.53+0.46+4.85−0.6013.4512.9610.0210.3613.0714.7513.5713.4712.19+4.66+3.70+5.22+5.34+1.48+4.35+4.86+4.87+3.72
PSNR↑Self-Calibrating GS* [3]180°18.5918.6817.4017.6718.4613.9019.39+4.91+4.39+4.34+6.45+6.39+5.98+7.0417.0117.4016.6116.4516.0118.1816.7919.7917.42+11.05+10.37+10.98+11.94+5.31+11.01+12.42+9.78+11.12
PSNR↑DirectFisheye-GS [32]180°23.1320.5620.5620.9322.9423.0723.35+1.09+2.24+1.76+4.70+2.76+0.62+2.3121.1220.5519.5119.9118.9221.8321.0622.2420.87+5.44+5.34+6.44+6.55+0.70+5.63+5.73+6.20+5.81
PSNR↑3DGRT [24]Full16.9714.8815.1715.8317.3414.4416.22+7.21+8.19+7.70+9.17+8.23+10.29+8.8822.8723.6321.1923.0621.3225.4322.8725.3724.10+2.66+2.32+3.03+2.84−0.32+2.03+2.35+2.17+2.29
PSNR↑3DGRT + PPISP [24, 4]Full19.1915.1117.8918.6919.1621.9716.92+4.99+7.96+4.98+6.31+6.41+2.76+8.1823.6723.7721.3823.0821.3725.6123.0825.4324.07+1.86+2.18+2.84+2.82−0.37+1.85+2.14+2.11+2.32
PSNR↑GOFish (Ours)Full24.1823.0722.8725.0025.5724.7325.1025.5325.9524.2225.9021.0027.4625.2227.5426.39
PSNR↑+ Geometry ObjectiveFull24.0822.8122.5624.9625.4624.4324.9425.2725.9324.2625.8919.6027.6225.2327.6826.27
SSIM↑Fisheye-GS* [22]145°0.7380.6100.6410.7120.6660.7060.729+0.055+0.096+0.103+0.116+0.211+0.093+0.1430.6430.6130.7190.5990.5790.6490.7170.6590.655+0.268+0.265+0.202+0.282+0.355+0.280+0.278+0.207+0.228
SSIM↑3DGUT [31]160°0.7910.6150.6730.7500.7960.8050.719+0.044+0.064+0.075+0.072+0.052−0.018+0.1050.8640.8240.8670.8100.7680.8250.8890.8850.855+0.017+0.027+0.018+0.015−0.038+0.024+0.012+0.019+0.020
SSIM↑3DGUT + PPISP [31, 4]160°0.7860.6280.6770.7620.7910.8340.721+0.055+0.063+0.072+0.071+0.057−0.017+0.1040.8660.8250.8670.8070.7670.8210.8880.8840.845+0.017+0.025+0.018+0.019−0.038+0.018+0.019+0.019+0.020
SSIM↑3DGEER* [14]178°0.7400.5650.5820.6880.7180.6460.795−0.005+0.066+0.076+0.023+0.000+0.021−0.0010.7380.7110.7310.6870.6210.7090.7620.7530.725+0.038+0.040+0.043+0.056+0.036+0.043+0.029+0.040+0.032
SSIM↑Self-Calibrating GS* [3]180°0.7370.5940.6310.7150.7400.7000.699+0.106+0.147+0.141+0.124+0.191+0.097+0.1700.7500.6700.7260.6040.4840.6790.6840.7710.695+0.206+0.216+0.187+0.268+0.293+0.260+0.246+0.159+0.195
SSIM↑DirectFisheye-GS [32]180°0.8130.6300.6800.7590.8110.7900.780+0.058+0.116+0.100+0.087+0.053+0.104+0.0770.8320.7980.8230.7730.7100.9010.8480.7550.800+0.055+0.059+0.068+0.064+0.018+0.071+0.046+0.159+0.195
SSIM↑3DGRT [24]Full0.7320.5610.6210.7350.7590.7570.647+0.123+0.173+0.160+0.129+0.110+0.074+0.2190.8850.8620.8900.8480.7890.9110.9140.8750.013+0.011+0.013+0.008−0.030+0.008+0.007+0.006+0.007
SSIM↑3DGRT + PPISP [24, 4]Full0.7540.5800.6690.7350.7760.7970.671+0.101+0.154+0.112+0.115+0.093+0.034+0.1950.8850.8640.8960.8490.7910.9110.9150.876+0.013+0.009+0.007+0.007−0.030+0.008+0.007+0.007+0.006
SSIM↑GOFish (Ours)Full0.8550.7340.7810.8500.8690.8310.8660.8980.9730.9030.8560.7590.9160.9220.882
SSIM↑+ Geometry ObjectiveFull0.8530.7180.7670.8490.8660.8270.8620.8950.9170.9030.8560.7590.9160.9220.880
LPIPS↓Fisheye-GS* [22]145°0.3610.4090.4020.4150.4690.4060.300+0.162+0.101+0.102+0.200+0.326+0.130+0.1510.4720.5200.4790.5040.5570.4840.4660.4860.490+0.377+0.403+0.396+0.430+0.485+0.340+0.415+0.369+0.354
LPIPS↓3DGUT [31]160°0.2790.3910.3600.3320.2990.2590.290+0.029+0.046+0.053+0.066+0.053−0.013+0.0840.2520.3110.2490.2460.2120.2980.2250.2750.285+0.036+0.046+0.039+0.021−0.036+0.044+0.026+0.051+0.040
LPIPS↓3DGUT + PPISP [31, 4]160°0.2840.3770.3560.3300.2970.2560.285+0.045+0.050+0.052+0.065+0.052−0.014+0.0810.2550.3100.2490.2510.2200.3000.2270.2570.285+0.073+0.079+0.087+0.061+0.018+0.072+0.061+0.080+0.081
LPIPS↓3DGEER* [14]178°0.4280.4180.4150.3200.2960.3730.203+0.001+0.052+0.055+0.035+0.014+0.031+0.0030.3320.3650.3430.3300.3140.3590.2920.3460.345+0.035+0.040+0.047+0.051+0.020+0.032+0.048+0.039+0.035
LPIPS↓Self-Calibrating GS* [3]180°0.3100.3910.3950.3600.3510.3690.285+0.123+0.121+0.117+0.161+0.159+0.093+0.1750.4900.4640.4230.4510.4570.4300.4180.4180.449+0.272+0.288+0.287+0.339+0.330+0.248+0.318+0.257+0.281
LPIPS↓DirectFisheye-GS [32]180°0.2560.3680.3510.3360.2950.2980.225+0.058+0.089+0.086+0.116+0.091+0.041+0.0770.2840.3330.2930.2750.2500.1980.3310.3020.304+0.073+0.079+0.086+0.061+0.018+0.072+0.061+0.080+0.081
LPIPS↓3DGRT [24]Full0.3050.4170.4020.3270.3030.3140.274+0.114+0.141+0.144+0.127+0.119+0.107+0.1770.2260.2670.2240.2120.1950.2530.1900.2330.250+0.025+0.026+0.024+0.064+0.005+0.017+0.061+0.021+0.024
LPIPS↓3DGRT + PPISP [24, 4]Full0.2810.3890.3530.3150.2870.2720.295+0.090+0.113+0.095+0.115+0.103+0.047+0.1580.2280.2630.2060.2020.1890.2350.2240.2300.241+0.073+0.079+0.086+0.061+0.018+0.072+0.061+0.080+0.081
LPIPS↓GOFish (Ours)Full0.1910.2760.2580.2000.1840.2250.1370.2010.2410.1870.1880.2000.2360.1740.2120.226
LPIPS↓+ Geometry ObjectiveFull0.1940.2880.2740.2050.1880.2310.1410.1960.3040.2950.2240.1740.2270.2590.2460.308

Table 8. Per-scene rendering results on FIORD [8] and FullCircle [5] for surface baselines. The best, second best, and third best entries are highlighted. Protocols follow Appendix § 13.§ 2.

MetricMethodFoVFIORD Full (200°)FIORD Common (90°)FullCircle Full (205°)
PSNR↑GOF [37]Rectified18.02 17.42 14.92 17.64 19.35 12.86 19.0330.96 24.15 22.44 28.19 31.88 30.60 29.2021.48 22.62 19.94 22.81 18.67 22.97 22.36 22.95 22.77
PSNR↑PGSR [2]Rectified17.82 17.12 14.56 18.36 18.81 12.59 18.4330.37 25.14 24.34 29.31 31.00 29.34 29.5421.08 23.35 20.04 23.30 18.31 23.08 22.46 23.10 23.59
PSNR↑SOF [26]Rectified18.08 16.92 13.26 17.00 19.30 12.73 19.0530.07 25.31 17.87 26.00 30.57 28.65 28.3221.13 22.41 19.55 22.72 19.62 22.03 22.02 22.67 23.63
PSNR↑MiLo [7]Rectified17.57 17.87 14.94 18.21 18.86 12.26 19.6730.02 25.18 24.35 29.46 30.25 29.19 29.0121.30 23.10 19.79 22.89 20.82 23.17 22.26 23.35 22.70
PSNR↑CoMe [27]Rectified18.18 19.34 14.91 18.64 19.59 12.77 20.0630.05 24.36 20.70 29.01 29.86 28.90 29.1122.00 22.95 20.18 22.90 19.94 23.50 22.31 23.34 23.50
PSNR↑GOFish (Ours)Full24.18 23.07 22.87 25.00 25.57 24.73 25.1030.61 26.61 25.24 30.94 30.72 27.86 31.0625.53 25.95 24.22 25.90 21.00 27.46 25.22 27.54 26.39
PSNR↑+ Geometry ObjectiveFull24.08 22.81 22.56 24.96 25.46 24.43 24.9430.57 25.94 24.67 30.63 30.29 26.99 30.6625.27 25.93 24.26 25.89 19.60 27.62 25.23 27.68 26.27
SSIM↑GOF [37]Rectified0.731 0.571 0.588 0.726 0.780 0.671 0.6640.936 0.773 0.805 0.901 0.945 0.927 0.8910.844 0.810 0.831 0.827 0.685 0.798 0.875 0.862 0.862
SSIM↑PGSR [2]Rectified0.715 0.582 0.584 0.737 0.760 0.662 0.6770.933 0.799 0.836 0.917 0.942 0.922 0.8980.828 0.807 0.825 0.783 0.649 0.792 0.862 0.855 0.822
SSIM↑SOF [26]Rectified0.741 0.543 0.496 0.705 0.789 0.674 0.6710.933 0.720 0.702 0.880 0.942 0.914 0.8770.837 0.807 0.822 0.788 0.696 0.786 0.877 0.859 0.835
SSIM↑MiLo [7]Rectified0.723 0.589 0.598 0.745 0.774 0.655 0.6770.929 0.783 0.827 0.914 0.932 0.909 0.8830.817 0.834 0.789 0.707 0.804 0.873 0.864 0.826
SSIM↑CoMe [27]Rectified0.744 0.606 0.572 0.758 0.792 0.695 0.7050.934 0.774 0.763 0.919 0.946 0.929 0.9010.848 0.832 0.786 0.694 0.807 0.877 0.866 0.830
SSIM↑GOFish (Ours)Full0.855 0.734 0.781 0.850 0.869 0.831 0.8660.937 0.839 0.872 0.934 0.944 0.922 0.9310.898 0.873 0.903 0.856 0.759 0.869 0.918 0.922 0.882
SSIM↑+ Geometry ObjectiveFull0.853 0.718 0.767 0.849 0.866 0.827 0.8620.935 0.826 0.864 0.932 0.942 0.917 0.9280.895 0.871 0.901 0.856 0.732 0.867 0.916 0.920 0.880
LPIPS↓GOF [37]Rectified0.306 0.405 0.448 0.358 0.286 0.383 0.2860.211 0.379 0.385 0.293 0.186 0.228 0.2970.259 0.246 0.308 0.227 0.259 0.246 0.308 0.227 0.297
LPIPS↓PGSR [2]Rectified0.326 0.397 0.430 0.357 0.324 0.402 0.3070.355 0.343 0.273 0.202 0.252 0.268 0.2770.324 0.257 0.283 0.324 0.240 0.308 0.299
LPIPS↓SOF [26]Rectified0.293 0.426 0.501 0.369 0.273 0.394 0.2840.213 0.434 0.495 0.319 0.187 0.250 0.2800.265 0.317 0.282 0.248 0.237 0.313 0.220 0.292 0.282
LPIPS↓MiLo [7]Rectified0.321 0.406 0.439 0.373 0.307 0.424 0.3080.220 0.376 0.364 0.302 0.220 0.289 0.3020.267 0.280 0.261 0.242 0.235 0.303 0.235 0.303 0.302
LPIPS↓CoMe [27]Rectified0.292 0.383 0.457 0.319 0.260 0.363 0.2520.209 0.378 0.435 0.263 0.163 0.215 0.2450.311 0.275 0.253 0.241 0.295 0.220 0.282 0.288
LPIPS↓GOFish (Ours)Full0.191 0.276 0.258 0.200 0.184 0.225 0.1370.193 0.289 0.281 0.218 0.167 0.220 0.1800.201 0.241 0.187 0.188 0.200 0.236 0.174 0.212 0.226
LPIPS↓+ Geometry ObjectiveFull0.194 0.288 0.274 0.205 0.188 0.231 0.1410.196 0.304 0.295 0.224 0.174 0.227 0.1840.245 0.192 0.189 0.235 0.241 0.174 0.216 0.225
MethodFoV0a5c8d56bb87d415e8eafe170a5c8d56bb87d415e8eafe17BarnCat.Court.Ign.Meet.Truck
Fisheye-GS* [22]145°24.5024.1729.2126.1329.2824.4028.3023.7522.5822.3525.9025.41
3DGUT [31]160°29.5426.0930.9728.2133.1025.6629.1023.6522.8422.8226.3025.96
3DGUT + PPISP [31, 4]160°29.8425.9430.9928.1133.1625.5430.1826.0024.0225.0127.0626.79
3DGEER* [14]178°27.5026.2931.2228.3732.1626.0128.6323.5022.9121.8425.2324.39
Self-Calibrating GS* [3]180°28.5025.7129.6026.3631.2124.6226.70*22.45*18.37*21.21*23.33*23.93*
DirectFisheye-GS [32]180°29.7026.6831.5228.4133.4025.9529.0024.4123.4222.7326.7425.91
3DGRT [24]Full29.1926.4030.9828.6231.9025.8927.6822.7822.0021.9524.4324.87
3DGRT + PPISP [24, 4]Full30.1726.7332.2928.9834.0526.0329.4526.0823.6825.1127.8626.35
GOF [37]Rectified28.1225.7130.0126.9229.4725.0430.4827.1132.4229.3334.2526.4526.2221.7720.2821.0124.7123.13
PGSR [2]Rectified28.1825.3229.5627.1929.3424.7730.4226.6231.8329.4833.6525.9627.8922.3020.8721.7224.0724.91
SOF [26]Rectified23.8125.6930.3226.9030.6925.1825.2126.7232.0528.6034.2026.1927.8222.5316.9321.7823.7025.00
MiLo [7]Rectified28.4125.7529.6627.4729.3425.0430.5327.0331.8729.4233.4626.1726.5621.4421.7620.6424.5722.80
CoMe [27]Rectified28.9726.0030.1927.8730.4425.2030.6127.0532.0829.5734.1226.0726.7921.1717.9921.4522.8624.42
GOFish (Ours)Full30.8426.9332.8129.4634.5926.3831.5127.4633.4030.0634.9426.9530.2526.3223.8124.9228.6526.65
+ Geometry ObjectiveFull30.7826.8832.7229.4434.5826.3331.4727.4233.3230.0334.9626.8830.2026.0423.3324.9428.0026.58
Fisheye-GS* [22]145°0.9050.8830.9260.8740.9300.8480.8590.8030.7970.8090.8830.874
3DGUT [31]160°0.9340.9000.9380.8910.9510.8670.8880.8240.8140.8110.8890.887
3DGUT + PPISP [31, 4]160°0.9330.9000.9380.8910.9510.8650.8920.8520.8250.8380.9060.896
3DGEER* [14]178°0.9290.9010.9390.8910.9520.8700.8670.8030.8010.7970.8830.871
Self-Calibrating GS* [3]180°0.9090.8850.9130.8380.9310.8310.804*0.731*0.659*0.725*0.818*0.814*
DirectFisheye-GS [32]180°0.9280.9020.9360.8810.9480.8640.8680.8090.8090.8140.8930.882
3DGRT [24]Full0.9200.8870.9300.8890.9440.8560.8760.8110.8030.8010.8780.881
3DGRT + PPISP [24, 4]Full0.9300.8980.9370.8910.9490.8610.8860.8470.8190.8330.9060.890
GOF [37]Rectified0.9260.8920.9340.8730.9460.8540.9370.9080.9460.9000.9570.8740.8820.8040.7520.8030.8760.885
PGSR [2]Rectified0.9240.8860.9260.8700.9400.8420.9360.9040.9390.9540.8620.8620.8900.8070.7820.8090.8700.880
SOF [26]Rectified0.8750.8950.9340.8730.9500.8590.9080.9090.9440.8920.9580.8730.8790.7990.6460.8080.8650.885
MiLo [7]Rectified0.9260.8930.9280.8750.9430.8490.9350.9080.9390.8970.9520.8650.8490.7560.7780.7700.8650.859
CoMe [27]Rectified0.9290.8980.9350.8890.9510.8530.9380.9110.9450.9060.9590.8660.8710.7660.6890.8080.8650.887
GOFish (Ours)Full0.9390.9070.9440.9030.9570.8770.9430.9150.9480.9040.9590.8830.8950.8570.8210.8490.9130.895
+ Geometry ObjectiveFull0.9390.9070.9440.9040.9570.8760.9430.9160.9480.9040.9590.8820.8930.8540.8120.8460.9090.893
Fisheye-GS* [22]145°0.2730.2700.2460.2150.2750.2580.2340.2650.2920.2010.2400.184
3DGUT [31]160°0.2070.2190.2130.1940.2220.2300.1940.2430.2620.2210.2410.166
3DGUT + PPISP [31, 4]160°0.2070.2160.2120.1930.2190.2300.1890.2120.2480.1890.2150.141
3DGEER* [14]178°0.2010.2040.2060.1750.2050.2130.2300.2490.2880.2070.2390.197
Self-Calibrating GS* [3]180°0.2290.2310.2380.2390.2450.2600.292*0.330*0.470*0.297*0.346*0.260*
DirectFisheye-GS [32]180°0.2010.2060.2100.1900.2150.2230.2190.2590.2720.2020.2310.178
3DGRT [24]Full0.3200.3300.2780.1930.2770.2720.2370.2490.2990.2250.2280.156
3DGRT + PPISP [24, 4]Full0.2720.2790.2450.1880.2460.2540.1870.2210.2530.1910.2060.140
GOF [37]Rectified0.2030.2220.2150.1820.2140.2350.1980.2200.2110.1760.2080.2340.2020.2450.3440.1960.2400.159
PGSR [2]Rectified0.2340.2560.2460.1990.2540.2550.2170.2460.2340.1920.2340.2520.1760.2400.2920.1880.2340.145
SOF [26]Rectified0.2950.2070.2140.1750.1980.2190.2620.2110.2120.1770.2000.2250.2000.2450.4590.1900.2380.157
MiLo [7]Rectified0.2260.2370.2380.2100.2360.2560.2220.2360.2310.2110.2280.2580.2550.3340.3240.2840.2720.216
CoMe [27]Rectified0.1750.1900.2060.1590.1830.2260.1760.1950.2050.1620.1860.2320.2090.2820.4260.1920.2210.138
GOFish (Ours)Full0.1570.1760.1800.1450.1800.1900.1710.1910.1870.1660.1880.2080.1790.1960.2570.1490.1980.140
+ Geometry ObjectiveFull0.1560.1740.1800.1450.1790.1900.1700.1890.1860.1660.1880.2080.1840.2030.2700.1540.2030.143

Table 9. Per-scene rendering results on ScanNet++ [34] and Tanks & Temples [18]. The best, second best, and third best entries are highlighted. Protocols follow Appendix § 13.§ 2.

MethodKit.RoadBld.Meet.Bri.Hall0a5c8d56bb87d415e8eafe17BarnCat.Court.Ign.Meet.Truck
Precision↑
GOF [37]0.4880.0160.1080.2920.0310.0740.5390.5460.6150.6470.5620.5660.5220.4020.4320.7260.4000.585
PGSR [2]0.6470.0250.1760.3460.0660.0900.7020.6430.6960.7860.6480.7150.5210.3370.1900.8310.3680.611
SOF [26]0.4730.0060.1030.3070.0310.0600.5610.5550.7020.6750.6080.6220.5860.3930.5060.6820.3990.553
MiLo [7]0.5320.0120.1370.3070.0270.0650.7050.5900.6190.7060.6440.6460.5510.3680.4130.7740.4260.602
CoMe [27]0.5080.0160.1480.3520.0390.0770.6390.5620.7160.6670.6800.6480.5580.4460.5430.8010.4910.604
GOFish (Ours)0.4780.0220.1420.4200.0740.0640.6620.5790.6620.6360.6700.6150.3630.2980.3890.4520.3710.461
+ Geometry Objective0.5070.0190.1350.4070.0760.0670.6470.5610.6710.6360.6510.6170.3950.3160.3890.5110.3730.454
Recall↑
GOF [37]0.6410.0360.1930.4930.0740.0540.6340.6260.5460.7510.5650.6510.4480.4000.2050.6260.2090.615
PGSR [2]0.6410.0970.1620.4810.3190.0550.6380.6010.6770.7450.6670.7330.5630.6860.3860.7930.3500.733
SOF [26]0.6470.0110.1280.5130.0930.0490.6300.6510.6310.7740.6150.7000.4850.4090.1980.6710.2150.554
MiLo [7]0.7130.0290.2010.5210.0630.0390.7540.6550.6290.7790.6640.7240.3920.3520.2380.6150.2040.574
CoMe [27]0.7090.0190.2180.6000.1700.0600.7490.7050.6690.8010.7360.7190.5070.4890.2450.7630.3010.665
GOFish (Ours)0.7280.0410.2440.6190.2830.0710.8770.7430.6940.8700.7810.8040.4470.3760.2900.6270.2640.527
+ Geometry Objective0.7640.0440.2380.6230.2950.0780.8720.7510.7290.8690.7460.8380.4700.4060.2770.6840.2610.514
F-score↑
GOF [37]0.5540.0220.1390.3670.0440.0630.5830.5830.5780.6950.5630.6060.4820.4010.2780.6720.2750.599
PGSR [2]0.6440.0400.1690.4030.1090.0680.6680.6210.6860.7650.6570.7240.5410.4520.2540.8110.3590.667
SOF [26]0.5470.0080.1140.3840.0460.0540.5930.5990.6640.7210.6110.6590.5300.4010.2850.6760.2800.553
MiLo [7]0.6090.0170.1630.3870.0380.0490.7290.6200.6240.7410.6540.6830.4580.3600.3020.6850.2760.588
CoMe [27]0.5920.0180.1760.4440.0640.0680.6890.6250.6920.7280.7070.6820.5310.4670.3380.7810.3730.633
GOFish (Ours)0.5770.0290.1800.5010.1170.0670.7540.6510.6770.7350.7210.6970.4010.3330.3320.5250.3080.492
+ Geometry Objective0.6100.0270.1720.4920.1210.0720.7430.6420.6990.7350.6950.7110.4290.3550.3240.5850.3070.482

Table 10. Per-scene geometry results. The best, second best, and third best entries are highlighted. Protocols follow Appendix § 13.§ 2.

Additional qualitative rendering results on FIORD

Figure 10. Additional qualitative rendering results on FIORD [8]. Metrics are per-scene.

Additional qualitative rendering and mesh results on FIORD

Figure 11. Additional qualitative rendering and mesh results on FIORD [8]. Metrics are per-scene.

Additional qualitative rendering results on FullCircle

Figure 12. Additional qualitative rendering results on FullCircle [5]. Metrics are per-scene.

Qualitative rendering results comparing Fisheye-GS, 3DGUT + PPISP, 3DGEER, Self-Cal. GS, DirectFisheye-GS, GOFish (Ours), and Ground Truth

Figure 13. Qualitative rendering results on ScanNet++ [34]. Metrics are per-scene.

Qualitative rendering and mesh results on ScanNet++

Figure 14. Qualitative rendering and mesh results on ScanNet++ [34]. Metrics are per-scene.

References

  1. [1]Pranav Asthana, Alex Hanson, Allen Tu, Tom Goldstein, Matthias Zwicker, and Amitabh Varshney. SplatSuRe: Selective super-resolution for multi-view consistent 3D Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11840–11849, 2026.arxiv.org/abs/2512.02172
  2. [2]Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. PGSR: Planar-based Gaussian Splatting for efficient and high-fidelity surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 31(9):6100–6111, 2025.DOI
  3. [3]Youming Deng, Wenqi Xian, Guandao Yang, Leonidas Guibas, Gordon Wetzstein, Steve Marschner, and Paul Debevec. Self-calibrating Gaussian Splatting for large field-of-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 25124–25133, 2025.arxiv.org/abs/2502.09563
  4. [4]Isaac Deutsch, Nicolas Moënne-Loccoz, Gavriel State, and Zan Gojcic. PPISP: Physically-plausible compensation and control of photometric variations in radiance field reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7289–7298, 2026.arxiv.org/abs/2601.18336
  5. [5]Yalda Foroutan, Ipek Oztas, Daniel Rebain, Aysegul Dundar, Kwang Moo Yi, Lily Goli, and Andrea Tagliasacchi. FullCircle: Effortless 3D reconstruction from casual 360° captures. arXiv preprint arXiv:2603.22572, 2026.arxiv.org/abs/2603.22572
  6. [6]Antoine Guédon and Vincent Lepetit. SuGaR: Surface-aligned Gaussian Splatting for efficient 3D mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354–5363, 2024.arxiv.org/abs/2311.12775
  7. [7]Antoine Guédon, Diego Gomez, Nissim Maruani, Bingchen Gong, George Drettakis, and Maks Ovsjanikov. MILo: Mesh-in-the-loop Gaussian Splatting for detailed and efficient surface reconstruction. ACM Transactions on Graphics, 44(6):1–15, 2025.
  8. [8]Ulas Gunes, Matias Turkulainen, Xuquian Ren, Arno Solin, Juho Kannala, and Esa Rahtu. FIORD: a fisheye indoor-outdoor dataset with LIDAR ground truth for 3D scene reconstruction and benchmarking. In Proceedings of the Scandinavian Conference on Image Analysis (SCIA), pages 3–17, 2025. 1, 2, 4, 5, 6, 7, 8, 13, 14, 17, 18, 19, 21, 22arxiv.org/abs/2504.01732
  9. [9]Alex Hanson, Allen Tu, Geng Lin, Vasu Singla, Matthias Zwicker, and Tom Goldstein. Speedy-Splat: Fast 3D Gaussian Splatting with sparse pixels and sparse primitives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21537–21546, 2025. 8arxiv.org/abs/2412.00578
  10. [10]Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Goldstein. PUP 3D-GS: Principled uncertainty pruning for 3D Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5949–5958, 2025. 8arxiv.org/abs/2406.10219
  11. [11]Yun He, Kelin Yu, and Matthias Zwicker. SceneOrchestra: Efficient agentic 3D scene synthesis via full tool-call trajectory generation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 521–538, 2026. 8arxiv.org/abs/2604.19907
  12. [12]Jan Held, Sanghyun Son, Renaud Vandeghen, Daniel Rebain, Matheus Gadelha, Yi Zhou, Anthony Cioppa, Ming C. Lin, Marc Van Droogenbroeck, and Andrea Tagliasacchi. MeshSplatting: Differentiable rendering with opaque meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7320–7329, 2026. 2arxiv.org/abs/2512.06818
  13. [13]Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for geometrically accurate radiance fields. In Proceedings of the ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 2DOI
  14. [14]Zixun Huang, Cho-Ying Wu, Yuliang Guo, Xinyu Huang, and Liu Ren. 3DGER: 3D Gaussian rendering made exact and efficient for generic cameras. In International Conference on Learning Representations (ICLR), 2026. 1, 2, 4, 16, 17, 19, 20arxiv.org/abs/2505.24053
  15. [15]Simon J. Julier and Jeffrey K. Uhlmann. New extension of the Kalman filter to nonlinear systems. In Signal Processing, Sensor Fusion, and Target Recognition VI, pages 182–193. SPIE, 1997. 2DOI
  16. [16]Sagi Katz, Ayellet Tal, and Ronen Basri. Direct visibility of point sets. ACM Transactions on Graphics, 26(3):24, 2007. 15
  17. [17]Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3D Gaussian Splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1–14, 2023. 1, 2, 4, 11, 12, 16, 17arxiv.org/abs/2308.04079
  18. [18]Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4):1–13, 2017. 2, 5, 7, 13, 15, 16, 17, 18, 20, 21, 26
  19. [19]Suyoung Lee, Jaeyoung Chung, Jaeyoo Huh, and Kyoung Mu Lee. OGDS: 3D scene reconstruction from omnidirectional images with 3D Gaussian Splattings. In Advances in Neural Information Processing Systems (NeurIPS), pages 57050–57075, 2024. 2arxiv.org/abs/2410.20686
  20. [20]Longwei Li, Xuhui Huang, Sai-Kit Yeung, and Hui Cheng. OmniGS: Fast radiance field reconstruction using omnidirectional Gaussian Splatting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2260–2268, 2025. 2 Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292–3310, 2023. 1
  21. [22]Zimu Liao, Siyan Chen, Rong Fu, Yi Wang, Zhongling Su, Hao Luo, Li Ma, Lanning Xu, Bo Dai, Hengjie Li, Zhilin Pei, and Xingcheng Zhang. Fisheye-GS: Lightweight and extensible Gaussian Splatting module for fisheye cameras. arXiv preprint arXiv:2409.04751, 2024. 1, 2, 4, 15, 16, 17, 19, 20arxiv.org/abs/2409.04751
  22. [23]Jiaqi Lin, Zihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. VastGaussian: Vast 3D Gaussians for large scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5166–5175, 2024. 2, 4arxiv.org/abs/2402.17427
  23. [24]Nicolas Moenne-Loccoz, Ashkan Mirzaei, OPerel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics, 43(6):1–19, 2024. 1, 2, 4, 7, 8, 15, 16, 17, 19, 20
  24. [25]Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. StopThePop: Sorted Gaussian Splatting for view-consistent real-time rendering. ACM Transactions on Graphics, 43(4):1–17, 2024. 3, 11arxiv.org/abs/2402.00525
  25. [26]Lukas Radl, Felix Windisch, Thomas Deixelberger, Jozef Hladky, Michael Steiner, Dieter Schmalstieg, and Markus Steinberger. SOF: Sorted opacity fields for fast unbounded surface reconstruction. In Proceedings of the SIGGRAPH Asia 2025 Conference Papers, pages 1–11, 2025. 1, 2, 4, 5, 6, 8, 11, 12, 15, 16, 17, 18, 19, 20, 21arxiv.org/abs/2506.19139
  26. [27]Lukas Radl, Felix Windisch, Andreas Kurz, Thomas Köhler, Michael Steiner, and Markus Steinberger. Confidence-based mesh extraction from 3D Gaussians. In Proceedings of the European Conference on Computer Vision (ECCV), pages 307–325, 2026. 1, 2, 5, 6, 12, 15, 16, 17, 19, 20, 21arxiv.org/abs/2603.24725
  27. [28]Gabriel Taubin. A signal processing approach to fair surface design. In Proceedings of the 22nd Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), pages 351–358, 1995. 8
  28. [29]Allen Tu, Haiyang Ying, Alex Hanson, Yonghan Lee, Tom Goldstein, and Matthias Zwicker. SpeedyDe3DGS: Speedy deformable 3D Gaussian Splatting with temporal pruning and motion grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26083–26093, 2026. 8arxiv.org/abs/2506.07917
  29. [30]Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. DN-Splatter: Depth and normal priors for Gaussian Splatting and meshing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2421–2431, 2025. 8arxiv.org/abs/2403.17822
  30. [31]Qi Wu, Janick Martinez Esturo, Ashkan Mirzaei, Nicolas Moënne-Loccoz, and Zan Gojcic. 3DGUT: Enabling distorted cameras and secondary rays in Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26036–26046, 2025. 1, 2, 3, 4, 5, 7, 8, 11, 15, 16, 17, 19, 20arxiv.org/abs/2412.12507
  31. [32]Zhengxian Yang, Fei Xie, Xutao Xue, Rui Zhang, Taicheng Huang, Yang Liu, Mengqi Ji, and Tao Yu. DirectFisheye-GS: Enabling native fisheye input in Gaussian Splatting with cross-view joint optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4942–4952, 2026. 1, 2, 4, 15, 16, 17, 19, 20arxiv.org/abs/2604.00648
  32. [33]Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. AbsGS: Recovering fine details in 3D Gaussian Splatting. In Proceedings of the 32nd ACM International Conference on Multimédia (MM), pages 1053–1061, 2024. 13DOI
  33. [34]Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A high-fidelity dataset of 3D indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12–22, 2023. 2, 4, 5, 6, 13, 15, 16, 17, 18, 20, 21, 24, 25
  34. [35]Senthil Yogamani, Ciarán Hughes, Jonathan Horgan, Ganesh Sistu, Padraig Varley, Derek O’Dea, Michal Uricar, Stefan Milz, Martin Simon, Karl Amende, et al. WoodScape: A multi-task, multi-camera fisheye dataset for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9308–9318, 2019. 1arxiv.org/abs/1905.01489
  35. [36]Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-Splatting: Alias-free 3D Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447–19456, 2024. 2, 3, 12arxiv.org/abs/2311.16493
  36. [37]Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 43(6):1–13, 2024. 1, 2, 4, 5, 8, 12, 13, 15, 16, 17, 18, 19, 20, 21arxiv.org/abs/2404.10772
  37. [38]Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586–595, 2018. 15arxiv.org/abs/1801.03924
  38. [39]Ziyu Zhang, Binbin Huang, Hanqing Jiang, Liyang Zhou, Xiaojun Xiang, and Shuhan Shen. Quadratic Gaussian Splatting: High quality surface reconstruction with second-order geometric primitives. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 28260–28270, 2025. 2arxiv.org/abs/2411.16392
  39. [40]Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus Gross. EWA splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223–238, 2002. 2DOI

Paper details

vlnc:W_3dv6vegh
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents