betaFeedback

WorksComputer Science

RIDE-Splat: Reflection Removal for In-the-Wild Driving Environments

Abstract

Driving video captured through a windshield violates the world-space consistency assumed by 3D Gaussian Splatting (3DGS): the road scene exhibits parallax while cabin reflections remain attached to the camera, causing standard 3DGS to embed reflected content into the reconstructed scene. We introduce RIDE-Splat, a pipeline for reflection-free 3D reconstruction from posed driving video without clean images, reflection masks, or pretrained networks. RIDE-Splat assigns world-consistent transmission through the windshield to a 3DGS scene and the dominant camera-attached reflection to one shared, signed RGB image. It builds the clean reconstruction in two passes. First, it learns the 3D geometry and the signed reflection image, representing both added radiance and transmission attenuation. Then, it uses the reflection-removed training images to fit a clean 3DGS model. At inference, the reflection image is discarded, preserving the standard 3DGS rendering path. We also introduce the RIDE-Splat Dataset, the first 3D reconstruction benchmark for reflection removal from driving video – with real driving scenes, physical windshield reflections in training views, and separately captured non-reflective evaluation images. Across the RIDE-Splat Dataset, XLD, and Wayve, RIDE-Splat outperforms 3DGS, DC-Gaussian, and RobustSplat on all nine aggregate metrics, improving PSNR by up to 3.59 dB and reducing LPIPS by up to 0.0876. We achieve the highest metrics in 38 out of 42 per-scene comparisons and removes windshield reflections that remain visible in the baselines while preserving the transmitted scene.

Introduction

3D Gaussian Splatting (3DGS) enables real-time, photorealistic novel-view synthesis [1], but its standard formulation assumes that posed images depict a coherent world-space scene. Driving video captured through a windshield violates this assumption because each frame contains the transmitted road scene and reflections of the vehicle interior. Camera motion reveals the key separation cue: the road scene exhibits parallax and remains consistent in world coordinates, while the dominant windshield reflection moves with the camera and remains approximately fixed in image coordinates. Standard 3DGS can therefore fit reflected content as scene geometry or appearance, contaminating novel views. Single-image reflection removal cannot exploit this multi-view cue, and rejecting contaminated observations discards valid transmission supervision. Conversely, a flexible per-view reflection field can mistakenly absorb road appearance that belongs in 3D. These failures motivate a constrained decomposition that assigns each signal to the coordinate system in which it remains consistent.

Historical dashcam archives and consumer driving videos document diverse roads, weather, lighting, and rare events that would be costly or impossible to recapture. Yet they remain underused for 3D reconstruction because windshield reflections make them poorly suited to pipelines designed for clean, purpose-captured imagery. Recovering reflection-free 3D scenes from these recordings could support closed-loop simulation, sensor-realistic augmentation, and rare-event replay without dedicated recapture.

We therefore introduce RIDE-Splat, a non-generative 3DGS formulation for reflection-free novel-view synthesis using only posed driving video with windshield reflections. It assigns world-consistent transmission through the windshield to a 3DGS scene and the dominant camera-attached reflection to one shared, signed RGB image. Because the reflection image is fixed in image coordinates and shared across views, it can represent image-stationary reflection content but cannot follow parallax-bearing road texture. Its signed values capture both added reflected radiance and transmission attenuation. The first pass uses a geometry warm-up to establish scene structure and a sparsity penalty to bias ambiguous content toward the 3D scene. Reflection self-localization then derives a fixed spatial weight from the recovered reflection image and refits both representations from scratch with the weight frozen. This second pass suppresses scene leakage without adding optimized parameters. At inference, we discard the reflection image, using standard 3DGS rasterization for rendering.

Existing benchmarks force a tradeoff: datasets with exact non-reflective targets use synthetic reflections, while those with physical windshield reflections lack clean targets. The RIDE-Splat Dataset closes this gap as the first 3D reconstruction benchmark designed specifically for reflection removal from driving video. For each real driving scene, it pairs a training traversal containing physical windshield reflections with a separately captured non-reflective evaluation traversal of the same route. The paired traversals are recorded within minutes using the same rigid camera mount, with only the reflection-inducing materials removed. All methods train only on the reflective traversal, with the non-reflective images held out as evaluation targets.

Across the RIDE-Splat Dataset, XLD [4], and Wayve [5], RIDE-Splat outperforms 3DGS, DC-Gaussian [2], and RobustSplat [3] on every aggregate metric under both physical and synthetically added reflections. It improves PSNR by up to 3.59 dB and LPIPS by up to 0.0876. These gains are consistent across scenes, with RIDE-Splat achieves higher performance in 38 of 42 per-scene metric comparisons. As shown in Fig. fig:teaser, RIDE-Splat removes windshield reflections that remain visible in the baselines while better preserving the transmitted scene.

In summary, we propose the following contributions:

  1. RIDE-Splat. A coordinate-based 3DGS method that separates world-consistent transmission from camera-attached windshield reflections.

  2. RIDE-Splat Dataset. The first 3D reconstruction benchmark for reflection removal from driving video.

Reflection removal has primarily been studied in image space. Single-image methods separate transmission and reflection using handcrafted or learned priors [6]. Multi-frame methods use layered motion, low-rank structure, or temporal consistency to distinguish the two signals [7, 8, 9]. RIDE-Splat instead brings this motion cue into 3D reconstruction, assigning world-consistent transmission to the 3D scene and the dominant windshield reflection to image coordinates.

Neural radiance fields [10] and 3D Gaussian Splatting (3DGS) [1] represent view-dependent radiance in world coordinates. Reflection-aware variants of these representations model glossy appearance, separate transmitted and reflected radiance, or reconstruct specular and semi-transparent surfaces [11, 12, 13, 14, 15]. Flash-Splat [16] separates transmission and reflection with Gaussian primitives but requires paired flash and no-flash captures. RIDE-Splat instead targets driving videos and learns a windshield-reflection layer fixed in image coordinates from reflective monocular video alone.

In-the-wild 3D reconstruction methods suppress transient or view-inconsistent distractors through sample rejection or transient modeling [17, 18, 19]. DF3DV-1K [20] benchmarks anti-distractor 3DGS methods [21, 22, 18, 23, 24, 25, 26], but does not include reflections of the vehicle interior on the windshield. RobustSplat [3], the strongest single-model baseline, delays Gaussian densification and bootstraps transient masks across scales. Our reflective weight maps are inspired by SplatSuRe [27] which derives per-pixel weight maps from multi-view sampling fidelity, but targets multi-view-consistent super-resolution rather than reflection removal. Driving-scene methods reconstruct large environments and dynamic actors [28, 29, 30, 31], but do not model windshield reflections.

DC-Gaussian [2] is the only prior 3D reconstruction method designed specifically to remove windshield reflections. It predicts a per-view reflection field from camera pose using a hash grid and MLP, while a latent code modulates the field’s intensity over time. RIDE-Splat instead uses one shared, signed image to impose a coordinate-based separation prior; this representation is orthogonal to DC-Gaussian’s geometry enhancement. In § 6, RIDE-Splat outperforms DC-Gaussian and RobustSplat across all metrics.

Method

Given NN reflective frames {Ii}i=1N\{I_i\}_{i=1}^{N} and corresponding camera poses {πi}i=1N\{\pi_i\}_{i=1}^{N}, RIDE-Splat reconstructs a 3D Gaussian Splatting (3DGS) representation [1] of the transmitted scene for reflection-free novel-view rendering, using no paired clean images or reflection masks. Figure 1 summarizes the two-stage pipeline. First, we jointly optimize the 3DGS scene GG and a single shared, signed reflection image RR. We then derive a fixed spatial weight WW from the recovered RR and refit the model from scratch. At inference, RR is discarded and only GG is rendered.

Figure 1. Overview of RIDE-Splat. Given posed driving frames with windshield reflections, RIDE-Splat reconstructs a reflection-free 3D Gaussian Splatting (3DGS) scene without paired clean images or reflection masks. Phase I: Coordinate-Based Joint Optimization (§ 3.2) warms up scene geometry, then jointly separates a world-consistent 3DGS scene from a shared, signed 2D reflection layer. Phase II: Reflection Self-Localization (§ 3.3) derives a continuous spatial weight from the recovered reflection layer and retrains the decomposition from scratch with this support fixed, preventing weak scene residuals from entering the final reconstruction. This refit adds no optimized parameters. At inference, the reflection layer is discarded and standard 3DGS rasterization renders reflection-free novel views.

3D Scene and Signed Reflection Model

Let G={gk}k=1KG=\{g_k\}_{k=1}^{K} denote a 3DGS scene with KK Gaussians. For camera πi\pi_i, rasterization produces the transmission image Ti=render⁡(G;πi)T_i=\operatorname{render}(G;\pi_i).

To motivate a signed reflection image, consider windshield image formation. At each pixel uu, α(u)∈[0,1]\alpha(u)\in[0,1] is an attenuation coefficient broadcast across RGB channels, and Lr(u)L_r(u) is the added RGB reflected radiance. The observed image is approximated by

Ii≈(1−α)⊙Ti+Lr=Ti+(Lr−α⊙Ti).\begin{equation} I_i \approx (1-\alpha)\odot T_i+L_r = T_i+\bigl(L_r-\alpha\odot T_i\bigr). \tag*{(1)} \end{equation}

Although LrL_r is nonnegative, the residual Lr−α⊙TiL_r-\alpha\odot T_i is generally signed and varies with TiT_i, capturing both added reflected radiance and attenuated transmission. RIDE-Splat approximates its dominant camera-static component with a single directly optimized, signed RGB image RR shared across all views. The predicted observed frame is

I^i=Ti+R≈Ii,i=1,…,N.\begin{equation} \hat I_i=T_i+R\approx I_i,\qquad i=1,\ldots,N. \tag*{(2)} \end{equation}

Because RR is shared across views, it can represent camera-attached reflection content but cannot follow scene parallax. The reflection model requires no view conditioning, network, hash grid, or per-frame coefficients.

Coordinate-Based Joint Optimization

For a world point XX, its pixel pi(X)=πi(X)p_i(X)=\pi_i(X) changes with camera motion, but the same 3D Gaussian can approximate its appearance in every view. For a rigidly mounted camera, the dominant cabin reflection exhibits little image-space motion. Encoding this reflection in the 3DGS scene GG would require inconsistent world geometry. Conversely, encoding parallax-bearing road texture in the fixed image RR would incur error across views. Joint optimization therefore favors scene content in GG and camera-static residuals in RR. The separation is not unique because RR can also absorb camera-static scene content such as the vehicle hood or a co-moving vehicle.

At iteration tt, a geometry warm-up gates the shared image, giving the prediction

I^i(t)=Ti+1[t>t0]R,\begin{equation} \hat I_i^{(t)}=T_i+\mathbbm{1}[t>t_0]R, \tag*{(3)} \end{equation}

where 1[t>t0]\mathbbm{1}[t>t_0] is an indicator function that starts optimization of RR after iteration t0t_0. We augment the standard 3DGS photometric loss with an ℓ1\ell_1 penalty on RR, weighted by β\beta:

L=Lphoto(I^i(t),Ii)+β∥R∥1.\begin{equation} \mathcal{L}=\mathcal{L}_{\text{photo}}(\hat I_i^{(t)},I_i)+\beta\lVert R\rVert_1. \tag*{(4)} \end{equation}

Averaging ∥⋅∥1\lVert\cdot\rVert_1 over channels and pixels keeps the reflection penalty on the same scale as the photometric objective. The regularizer resolves ambiguous pixels in favor of the 3DGS scene, while the warm-up establishes initial scene geometry before RR can absorb persistent image-space residuals.

Figure 2. Reflective training views from the Leesburg, Martin, and Reston scenes in the RIDE-Splat Dataset, shown from left to right. Physical windshield reflections are produced by common cabin objects, including newspapers, aluminum foil, and clips.

Reflection Self-Localization

Although sharing and sparsity constrain RR, coordinate-based joint optimization can still leave low-amplitude scene residuals in RR outside reflection regions. To prevent this leakage in the final decomposition, reflection self-localization estimates a continuous spatial weight WW from the recovered RR, restricting where the reflection model may act. For each pixel uu, we define the channel-max response a(u)=max⁡q∈{1,2,3}∣Rq(u)∣a(u)=\max_{q\in\{1,2,3\}}|R_q(u)|. The absolute value measures reflection magnitude regardless of sign, while the channel maximum retains any strong RGB response without dilution by weaker channels. We smooth aa with a Gaussian kernel κσ\kappa_\sigma of standard deviation σ\sigma to aggregate pixel-level responses into spatially coherent support, then normalize over all pixel locations vv:

b(u)=(κσ∗a)(u),W(u)=b(u)−min⁡vb(v)max⁡vb(v)−min⁡vb(v)+ϵ.\begin{equation} \begin{aligned} b(u) &= (\kappa_\sigma\ast a)(u), \\ W(u) &= \frac{b(u)-\min_v b(v)}{\max_v b(v)-\min_v b(v)+\epsilon}. \end{aligned} \tag*{(5)} \end{equation}

Here ϵ>0\epsilon>0 ensures numerical stability, the normalization yields a continuous spatial weight W(u)∈[0,1]W(u){\in}[0,1], and the bandwidth σ\sigma controls the extent of the estimated support. The scalar map WW determines where the signed RGB appearance RR can contribute. Because WW is inferred only after joint optimization, it is unavailable at initialization. We therefore restart the joint optimization of GG and RR from scratch with WW frozen, retaining the objective in Eq. (4) and the gated 3DGS warm-up. This trains the final decomposition under localized reflection support from the outset rather than inheriting leakage from the joint fit. Under this weighting, Eq. (3) becomes

Reff=R⊙W,I^i(t)=Ti+1[t>t0]Reff.\begin{equation} R_{\mathrm{eff}}=R\odot W,\qquad \hat I_i^{(t)}=T_i+\mathbbm{1}[t>t_0]R_{\mathrm{eff}}. \tag*{(6)} \end{equation}

The photometric term uses the weighted reflection, while the ℓ1\ell_1 penalty remains on RR. In regions with little reflection evidence, WW suppresses the effective reflection, forcing those pixels to be explained by GG. The component ablation in § 6.1 confirms that this self-localized refit provides additional gains over the shared reflection layer alone. Because WW is frozen, the second pass introduces no additional optimized parameters. At inference, we omit RR and render reflection-free novel views from GG using the standard 3DGS rendering path. We ablate our choice of β\beta and σ\sigma in § 6.2.

RIDE-Splat Dataset

Existing driving-scene 3D reconstruction benchmarks either retain exact clean targets by adding synthetic reflections or capture physical reflections without non-reflective targets. Neither setting directly measures transmission recovery under physical windshield reflections. To fill this gap, we introduce the RIDE-Splat Dataset, the first 3D reconstruction benchmark designed specifically for reflection removal from driving video. It pairs reflective training traversals of real driving scenes with separately captured non-reflective evaluation traversals, enabling quantitative evaluation against actual non-reflective images. Table tab:dataset_comparison compares the evaluation datasets.

We recorded H.264 clips in clear daytime conditions using a Garmin S35 dashcam at 1920×10801920{\times}1080 and 29.97 FPS. To diversify reflection appearance, we placed common cabin objects made of paper, laminated paper, aluminum, plastic, painted wood, and steel on the dashboard. The object set differs by scene but remains fixed within each reflective traversal, producing scene-specific, approximately camera-static reflections. Figure 2 shows representative reflective training views from all three scenes. To prepare the footage for 3D reconstruction, we removed the 40-pixel-tall dashcam metadata banner, undistorted and cropped the frames to 1900×10251900{\times}1025, and downsampled them to reduce dataset size. Table tab:ride_splits reports the downsampled frame rate and number of training and evaluation frames for each scene.

[]

DatasetDataReflectionsReference
XLD [4]SyntheticSyntheticExact
Wayve [5]RealSyntheticExact
RIDE-Splat Dataset (Ours)RealRealApproximate

Table 1. Evaluation datasets and RIDE-Splat splits.

[]

Number of Frames
3-4 SceneFPSReflective TrainingNon-Reflective Evaluation
Leesburg813430
Martin68230
Reston85230
Total–26890

The evaluated benchmark contains three scenes, Leesburg, Martin, and Reston, spanning a historic town center, a tree-lined institutional district, and suburban residential streets. These settings vary in traffic, building density, tree cover, and lateral parallax. For each scene, we recorded a reflective training traversal and a non-reflective reference traversal of the same route within minutes of one another. The camera remained in the same rigid mount, and we removed only the reflection-inducing materials between traversals. To reduce cross-traversal confounds, we visually selected sequences with few transient objects and minimal visible change. We jointly registered both traversals with COLMAP [32], placing their camera poses in a common coordinate frame. All geometry and camera poses were recovered from RGB footage alone, without LiDAR, odometry, depth, or other vehicle sensors.

Only reflective frames are used for training, without reflection masks or annotations, so methods must infer contaminated regions directly. Evaluation frames are uniformly sampled from each non-reflective traversal. Each method renders at these held-out camera poses and is scored against the corresponding non-reflective images.

The fixed camera mount, short capture interval, and sequence selection reduce cross-traversal variation, but traffic, illumination, and camera trajectories cannot be matched exactly. The resulting references are therefore approximate. Nevertheless, they contain minimal windshield reflections and are never used during optimization, so they directly evaluate transmission recovery under physical reflections. We will publicly release the RIDE-Splat Dataset, including the evaluated benchmark and additional captured scenes.

RIDE-Splat Dataset (Ours)XLDWayve
MethodPSNR↑\uparrowSSIM↑\uparrowLPIPS↓\downarrowPSNR↑\uparrowSSIM↑\uparrowLPIPS↓\downarrowPSNR↑\uparrowSSIM↑\uparrowLPIPS↓\downarrow
3DGS17.370.72040.544521.370.88240.488220.140.73660.4426
DC-Gaussian14.470.68210.554918.320.84360.527620.070.71610.4694
+ Oracle GT Color Correction20.250.74420.510923.070.88940.492020.490.71550.4713
RobustSplat18.300.73170.521621.640.88150.489120.190.73640.4424
RIDE-Splat (Ours)18.480.74280.487325.230.89800.400620.840.74220.4179

Table 2. Quantitative results on the RIDE-Splat Dataset, XLD [4], and Wayve [5]. Metrics are averaged across the evaluated scenes in each dataset, and the best result is bolded. Because DC-Gaussian exhibits substantial underexposure and color shifts, we also report a + Oracle GT Color Correction variant that applies the released color transform using each test target as reference. This variant deliberately favors DC-Gaussian and is excluded from bolding. Even so, RIDE-Splat achieves the strongest overall results.

Experiments

We evaluate RIDE-Splat on fully synthetic scenes, real driving imagery with synthetic reflections, and real driving imagery with physical windshield reflections. Together, these settings test whether it consistently recovers the transmitted scene as both scene content and reflection formation become more realistic. All methods receive identical reflective training images and camera poses, while non-reflective targets are used only for evaluation. Beyond the main comparison, we isolate the contributions of the Reflection Layer and self-localization, evaluate hyperparameter sensitivity, and stress-test the camera-static reflection prior under increasing drift.

Datasets

We evaluate on three datasets that vary both scene content and reflection formation. The RIDE-Splat Dataset pairs real driving scenes with physical windshield reflections. We use the Leesburg, Martin, and Reston splits in Table 1. XLD [4] pairs synthetic scenes with synthetic reflections, while Wayve [5] combines real driving imagery with synthetic reflections.

For XLD, we evaluate all six scenes: Town01, Town02, Town03, Town04, Town05, and Town10. We add a camera-static synthetic windshield reflection to 150 training images per town and evaluate all 28 clean views captured at a 1 m left lateral offset. For Wayve, we use the first five 10-second sequences, scene_001–scene_005, and add static synthetic reflections to 176 training images per sequence. Every eighth original frame is held out, yielding 24 clean evaluation images per sequence. Appendix § 9.1 details the synthetic-reflection construction and parameters for both datasets.

Baselines and Metrics

We use the official implementations and published settings for all baselines. Every method receives identical training images and camera poses and is evaluated against identical targets. We compare against 3DGS [1], DC-Gaussian [2], and RobustSplat [3]. DC-Gaussian specifically targets driving video with windshield reflections, while RobustSplat is the strongest single-model anti-distractor method in DF3DV-1K [20]. We report full-image PSNR, SSIM, and LPIPS.

DC-Gaussian Evaluation Protocol

DC-Gaussian [2] introduces DCVR, a self-captured real-world dataset with strong reflections, and also evaluates on BDD100K [33]. Both are real-world dashcam datasets with windshield obstructions, but neither provides non-reflective transmission targets. DC-Gaussian decomposes each view into a transmission image and an image-space obstruction image, then combines them using a learned opacity map. Its published PSNR, SSIM, and LPIPS compare this recomposed output with held-out observed frames. These metrics therefore measure whether the two predicted layers reproduce the observed obstructed frame rather than whether the transmission recovers the reflection-free scene. Our metrics instead compare transmission renders directly with non-reflective targets: exact clean images for XLD and Wayve, and separately captured images for the RIDE-Splat Dataset.

On real-world data, DC-Gaussian transmission renders exhibit substantial underexposure and color shifts. We report the unmodified renders as DC-Gaussian. To prevent these errors from dominating the comparison, the + Oracle GT Color Correction variant uses the released routine to fit one global quadratic color transform from each transmission render to its test target over all unsaturated pixel correspondences. This transform is not part of the published method or default evaluation and could be applied to any renderer. Because it uses test-target oracle information, the variant deliberately favors DC-Gaussian and is excluded from best-result highlighting.

Implementation Details

RIDE-Splat uses the original 3DGS optimization settings for GG. We optimize RR with Adam at a learning rate ηR=0.01\eta_R=0.01 and set β=0.01\beta=0.01. Following RobustSplat [3], we delay activation of the reflection image behind a geometry warm-up, ending at t0=2000t_0=2000 iterations. This places activation after densification begins at iteration 500, and well before it concludes, thus scene geometry is established before RR can compete for it. The self-localization pass uses σ=8\sigma=8 and repeats the same schedule with WW fixed. We ablate ηR\eta_R, β\beta, and σ\sigma in § 6.2. No pretrained image generator or reflection-removal network is used.

Results

Figure 3. Qualitative results on the RIDE-Splat Dataset, XLD [4], and Wayve [5]. The examples show Martin and Leesburg from the RIDE-Splat Dataset, Town02 from XLD, and scene_003 from Wayve. Compared with 3DGS [1], DC-Gaussian [2], and RobustSplat [3], RIDE-Splat more effectively removes windshield reflections while preserving the transmitted scene. DC-Gaussian exhibits visible color shifts, so Table 2 also reports the + Oracle GT Color Correction variant described in § 5.3 that uses each test target as reference.

Table 2 evaluates transmission quality against held-out non-reflective targets. The RIDE-Splat Dataset uses separately captured, approximately aligned non-reflective traversals, whereas XLD and Wayve use the original clean frames before reflections are added to the training images. Except for the explicitly marked color-corrected variant, test targets are used only for evaluation. Among methods without test-target access, RIDE-Splat ranks first on all nine aggregate metrics across physical and synthetically added reflections. Its largest margins occur on XLD, where it improves PSNR by 3.59 dB and reduces LPIPS by 0.0876 relative to the next-best baseline. Appendix § 9.2 reports per-scene results. Excluding the target-aware oracle correction, RIDE-Splat outperforms the baselines on 38 of 42 scene-metric comparisons, confirming that the aggregate gains are not driven by a small subset of scenes.

Because DC-Gaussian exhibits substantial underexposure and color shifts, Table 2 also reports the + Oracle GT Color Correction variant described in § 5.3. Even with this test-target access, DC-Gaussian remains behind RIDE-Splat on every metric for XLD and Wayve. On the RIDE-Splat Dataset, the corrected variant leads in PSNR and SSIM, while RIDE-Splat retains the best LPIPS. These target-aware scores are excluded from best-result highlighting.

Figure 3 presents representative comparisons from all three datasets. Vanilla 3DGS embeds windshield reflections into the reconstruction, while DC-Gaussian leaves residual reflections and visible color shifts. RobustSplat introduces broader image artifacts on real scenes. RIDE-Splat more effectively removes the reflections while preserving the transmitted road scene, yielding sharper building and road-sign textures than DC-Gaussian and fewer artifacts than RobustSplat. These visual gains confirm that the metric improvements reflect cleaner transmission rather than simply a global appearance change.

Component Ablation

Table 3 applies the method components cumulatively across XLD. Our coordinate-based joint-optimization method, described in § 3.2, provides the primary improvement across all metrics. Reflection self-localization, described in § 3.3, further improves all metrics without adding optimized parameters. The ablation confirms that coordinate-based separation drives the gain and that self-localization further refines the reconstructed scene.

Hyperparameter Sensitivity

We vary one hyperparameter at a time around the defaults ηR=0.01\eta_R=0.01, β=0.01\beta=0.01, and σ=8\sigma=8. Table 4 reports the results on Town01. Performance remains stable across the tested learning rates. Reducing β\beta to 0.002 increases the mean recovered ∣R∣\lvert R\rvert by 40% while leaving PSNR and SSIM nearly unchanged, suggesting that sharing the reflection image across views, rather than sparsity alone, prevents it from absorbing scene content. Increasing β\beta to 0.05 over-regularizes the reflection image and degrades all metrics. Among the tested smoothing scales, σ=32\sigma=32 yields the best LPIPS, whereas the narrower support at σ=2\sigma=2 lowers PSNR.

Robustness to Reflection Motion

We stress-test the camera-static assumption on Town01 with a single seed using the Reflection Layer from the Coordinate-Based Joint Optimization row from Table 3. Within a fixed support, the reflection appearance LrL_r is translated horizontally by the specified number of pixels per frame over 150 frames. At the largest drift of 1.2 pixels per frame, the cumulative displacement reaches 180 pixels, or 9.4% of the image width. Across the sweep, Reflection Layer PSNR varies by only 0.65 dB, compared with 1.45 dB for DC-Gaussian. The shared reflection model therefore tolerates moderate horizontal drift, although reflections with independent parallax or deformation remain outside its scope.

ConfigurationPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
3DGS [1]21.370.88240.4882
+ Joint Optimization24.780.89680.3994
+ Self-Localization25.230.89800.4006

Table 3. Component ablation on XLD [4]. Coordinate-based Joint Optimization (§ 3.2) and Reflection Self-Localization (§ 3.3) are applied cumulatively, metrics are averaged across all six towns, and the best value in each column is bolded.

ηR\eta_Rβ\betaσ\sigmaPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
0.010.01825.250.92550.4586
0.0050.01825.100.92640.4572
0.050.01825.340.92630.4640
0.010.002825.240.92690.4728
0.010.05822.690.91970.5045
0.010.01224.820.92350.4590
0.010.013225.320.92490.4444

Table 4. Hyperparameter sensitivity on Town01 from XLD [4]. Each row varies one default value with the others held fixed. The default row is the run reported in Table 2.

PSNR ↑\uparrow
2-3 Drift (Pixels)Reflection Layer (Ours)DC-Gaussian
0.024.7812.97
0.1524.7512.04
0.324.6213.02
0.524.8412.21
0.824.4613.11
1.224.1913.49
Range0.651.45

Table 5. PSNR on Town01 from XLD [4] under increasing reflection drift. The reflection appearance is translated horizontally within a fixed support by the stated number of pixels per frame over 150 frames. Range is the difference between the minimum and maximum PSNR across the tested drift levels.

Limitations

RIDE-Splat assumes that the dominant windshield reflection is semi-transparent and approximately fixed in image coordinates. The motion stress test indicates tolerance to moderate drift, but substantial independent motion or deformation may leave residual contamination. As a 3DGS-based method, it also inherits challenges under sparse forward-facing views, although its reflection formulation is orthogonal to driving-scene extensions such as Street Gaussians [28] and can be combined with them for large, dynamic environments. Evaluating real windshield reflections against non-reflective targets requires separate traversals, making exact temporal and environmental matching impossible and leaving the references approximate. Our physical-reflection evaluation currently covers a three-scene subset of the broader RIDE-Splat Dataset, and additional captured scenes will be publicly released.

Applications and Future Work

RIDE-Splat can convert existing dashcam video into reflection-free 3D assets for closed-loop simulation and rare-event replay. Beyond driving, its camera-coordinate prior may extend to robot cameras behind protective windows or transparent housings, where enclosure reflections move with the camera rather than the environment. Future work can relax the shared-image assumption through constrained low-dimensional reflection motion and adapt the self-localization weight using multi-view consistency, without introducing unrestricted per-view capacity that can absorb scene content. Expanding the RIDE-Splat Dataset across weather, lighting, camera systems, and reflection dynamics, together with more synchronized reference capture, will broaden evaluation and improve reference fidelity.

Conclusion

RIDE-Splat reframes windshield reflection removal as a coordinate-separation problem for 3D reconstruction from posed driving video. It assigns world-consistent transmission through the windshield to a 3D Gaussian Splatting (3DGS) [1] scene and the dominant camera-attached reflection to a shared, signed image, using camera motion to separate them without clean supervision. Reflection self-localization further restricts where the reflection layer can act in a second fit without adding optimized parameters, and discarding the layer at inference preserves the standard 3DGS rendering path. The RIDE-Splat Dataset provides the first 3D reconstruction benchmark for reflection removal from driving video, pairing training traversals containing physical windshield reflections with held-out non-reflective captures. Across the RIDE-Splat Dataset, XLD [4], and Wayve [5], RIDE-Splat outperforms 3DGS, DC-Gaussian [2], and RobustSplat [3] on every aggregate metric; per-scene and qualitative comparisons confirm that this advantage reflects cleaner transmission rather than a global appearance correction. Component ablations identify the shared reflection image as the primary source of improvement and self-localization as a consistent refinement. More broadly, modeling structured contamination in the coordinate system where it remains stable provides a simple principle for reconstructing scenes from driving archives and other video captured through glass.

Acknowledgments

This research is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 140D0423C0076. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Commercial support was provided by the Amazon Research Awards program and Open Philanthropy. Further support was provided by the NSF TRAILS Institute (2229885).

Synthetic Reflection Construction

Table 6 reports the synthesis parameters for XLD and Wayve. Both settings use static reflections with zero flicker and image-space drift. For Wayve, reflection support α\alpha is derived from a real windshield annotation, covers 31.8% of the image, and is shared across scene_001–scene_005. The support is used only to synthesize reflective training images and is never provided to a method. The added reflection LrL_r is a synthetic low-frequency wash with strength 0.50.5–0.70.7 for XLD and 0.50.5 for Wayve. Wayve also uses no center glare and lower obstruction strength, producing milder contamination that may contribute to the smaller PSNR gap between methods.

ParameterXLDWayve
Flicker0.00.0
Drift [px]00
Center glare0.350.0
Feather8.06.0
Obstruction blur14.012.0
Obstruction strength0.60.4
Side fraction0.220.0
LrL_r strength0.5–0.70.5

Table 6. Synthetic reflection parameters. XLD [4] and Wayve [5] use the same compositing pipeline without temporal variation. A dash indicates that no separate value is specified.

Per-Scene Metrics

Table 7, Table 8, and Table 9 report scene-level performance for the three RIDE-Splat Dataset scenes, all six XLD towns, and five Wayve sequences. They use the same methods and evaluation protocols as the main comparison in Table 2. Excluding the test-target correction, RIDE-Splat achieves the best PSNR on 12 of 14 scenes and the best SSIM and LPIPS on 13 of 14, leading 38 of 42 scene-metric comparisons. RobustSplat leads PSNR on Reston, DC-Gaussian leads PSNR on scene_002, and 3DGS leads SSIM and LPIPS on scene_001.

RIDE-Splat Dataset (Ours)XLDWayve
MethodLeesburgMartinRestonMeanTown01Town02Town03Town04Town05Town10Means001s002s003s004s005Mean
3DGS15.5518.6417.9117.3720.9720.3822.1321.6322.0721.0321.3719.9719.8420.1620.5820.1620.14
RobustSplat15.3720.1019.4218.3021.2220.7622.2121.9122.5221.2221.6419.8620.2320.1820.5720.1020.19
DC-Gaussian13.7315.6714.0114.4712.3620.4319.4715.7820.2221.6318.3219.8720.6419.3820.3920.0720.07
+ Oracle GT Color Correction17.1022.8420.8020.2522.6921.1523.5624.4824.0022.5523.0720.1220.7020.9920.5220.1120.49
RIDE-Splat (Ours)16.0320.1819.2118.4825.2524.7226.0026.0725.3124.0225.2320.1020.5420.6621.7421.1520.84

Table 7. PSNR on each scene in the RIDE-Splat Dataset, XLD [4], and Wayve [5]. The best result for each scene is bolded. The DC-Gaussian + Oracle GT Color Correction variant is excluded from bolding because it uses test-target information.

RIDE-Splat Dataset (Ours)XLDWayve
MethodLeesburgMartinRestonMeanTown01Town02Town03Town04Town05Town10Means001s002s003s004s005Mean
3DGS0.59400.81360.75380.72040.90770.87340.90800.92550.89170.78780.88240.79960.73250.71880.69260.73960.7366
RobustSplat0.59400.83440.76690.73170.91020.87480.90690.92150.89200.78380.88150.79590.73750.72010.69000.73870.7364
DC-Gaussian0.54980.79230.70400.68210.73880.87030.89900.87110.89150.79090.84360.77900.73000.70010.65370.71800.7161
+ Oracle GT Color Correction0.60850.84460.77970.74420.91430.87850.91360.93340.90090.79590.88940.77690.72890.70390.64990.71780.7155
RIDE-Splat (Ours)0.60700.84200.77940.74280.92550.89310.91880.94000.90200.80840.89800.79500.73960.72510.70220.74910.7422

Table 8. SSIM on each scene in the RIDE-Splat Dataset, XLD [4], and Wayve [5]. The best result for each scene is bolded. The DC-Gaussian + Oracle GT Color Correction variant is excluded from bolding because it uses test-target information.

RIDE-Splat Dataset (Ours)XLDWayve
MethodLeesburgMartinRestonMeanTown01Town02Town03Town04Town05Town10Means001s002s003s004s005Mean
3DGS0.60570.45400.57370.54450.53410.48270.41120.50610.45880.53640.48820.33210.46650.50590.46260.44610.4426
RobustSplat0.60070.43840.52570.52160.52960.48310.41530.50160.45670.54860.48910.34000.45750.50040.46760.44640.4424
DC-Gaussian0.60110.47990.58380.55490.61670.49210.42430.54180.53880.55200.52760.36890.47150.53580.49210.47870.4694
+ Oracle GT Color Correction0.59510.42310.51460.51090.56850.49220.41950.47870.44560.54780.49200.37340.47680.53390.49620.47620.4713
RIDE-Splat (Ours)0.57120.40250.48800.48730.45860.37760.34290.39200.37400.45840.40060.33760.43320.47000.43880.41000.4179

Table 9. LPIPS on each scene in the RIDE-Splat Dataset, XLD [4], and Wayve [5]. The best result for each scene is bolded. The DC-Gaussian + Oracle GT Color Correction variant is excluded from bolding because it uses test-target information.

References

  1. [1]B. Kerbl, G. Kopanas, T. Leimkuehler, and G. Drettakis, `` 3D Gaussian Splatting for Real-Time Radiance Field Rendering ,'' ACM Transactions on Graphics, vol. 42, no. 4, pp. 1--14, 2023.
  2. [2]L. Wang, K. Cheng, S. Lei, S. Wang, W. Yin, C. Lei, X. Long, and C.-T. Lu, `` DC-Gaussian: Improving 3D Gaussian Splatting for Reflective Dash Cam Videos ,'' in Advances in Neural Information Processing Systems 37, 2024, pp. 99\,898--99\,920.
  3. [3]C. Fu, Y. Zhang, K. Yao, G. Chen, Y. Xiong, C. Huang, S. Cui, and X. Cao, `` RobustSplat: Decoupling Densification and Dynamics for Transient-Free 3DGS ,'' in 2025 IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 27\,126--27\,136.
  4. [4]H. Li, C. Wu, M. Yuan, Y. Zhang, C. Zhao, C. Song, H. Feng, E. Ding, D. Zhang, and J. Wang, `` XLD: A Cross-Lane Dataset for Benchmarking Novel Driving View Synthesis ,'' in 2025 International Conference on 3D Vision (3DV), 2025, pp. 659--669.
  5. [5]J. Zürn, P. Gladkov, S. Dudas, F. Cotter, S. Toteva, J. Shotton, V. Simaiaki, and N. Mohan, ``Wayvescenes101: A dataset and benchmark for novel view synthesis in autonomous driving,'' 2024.
  6. [6]R. Wan, B. Shi, L.-Y. Duan, A.-H. Tan, and A. C. Kot, `` Benchmarking Single-Image Reflection Removal Algorithms ,'' in 2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 3942--3950.
  7. [7]R. Szeliski, S. Avidan, and P. Anandan, `` Layer extraction from multiple images containing reflections and transparency ,'' in Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No.PR00662), vol. 1, 2000, pp. 246--253.
  8. [8]A. Nandoriya, M. Elgharib, C. Kim, M. Hefeeda, and W. Matusik, `` Video Reflection Removal Through Spatio-Temporal Optimization ,'' in 2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2430--2438.
  9. [9]Y.-L. Liu, W.-S. Lai, M.-H. Yang, Y.-Y. Chuang, and J.-B. Huang, `` Learning to See Through Obstructions ,'' in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 14\,203--14\,212.
  10. [10]B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, `` NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis ,'' in Computer Vision -- ECCV 2020, 2020, pp. 405--421.
  11. [11]D. Verbin, P. Hedman, B. Mildenhall, T. Zickler, J. T. Barron, and P. P. Srinivasan, `` Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields ,'' in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 5481--5490.
  12. [12]Y. Jiang, J. Tu, Y. Liu, X. Gao, X. Long, W. Wang, and Y. Ma, `` GaussianShader: 3D Gaussian Splatting with Shading Functions for Reflective Surfaces ,'' in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 5322--5332.
  13. [13]Y. Yao, Z. Zeng, C. Gu, X. Zhu, and L. Zhang, `` Reflective Gaussian Splatting ,'' in The Thirteenth International Conference on Learning Representations, 2025.
  14. [14]Y.-C. Guo, D. Kang, L. Bao, Y. He, and S.-H. Zhang, `` NeRFReN: Neural Radiance Fields with Reflections ,'' in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 18\,388--18\,397.
  15. [15]J. Shi, X. Ying, B. Xing, R. Guo, and W. Yue, `` RT-Splatting: Joint Reflection-Transmission Modeling with Gaussian Splatting ,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026.
  16. [16]M. Xie, H. Cai, S. Shah, Y. Xu, B. Y. Feng, J.-B. Huang, and C. A. Metzler, `` Flash-Splat: 3D Reflection Removal with Flash Cues and Gaussian Splats ,'' in Computer Vision -- ECCV 2024, 2024, pp. 122--139.
  17. [17]S. Sabour, S. Vora, D. Duckworth, I. Krasin, D. J. Fleet, and A. Tagliasacchi, `` RobustNeRF: Ignoring Distractors with Robust Losses ,'' in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 20\,626--20\,636.
  18. [18]S. Sabour, L. Goli, G. Kopanas, M. Matthews, D. Lagun, L. Guibas, A. Jacobson, D. Fleet, and A. Tagliasacchi, `` SpotLessSplats: Ignoring Distractors in 3D Gaussian Splatting ,'' ACM Transactions on Graphics, vol. 44, no. 2, pp. 1--11, 2025.
  19. [19]J. Lin, J. Gu, L. Fan, B. Wu, Y. Lou, R. Chen, L. Liu, and J. Ye, `` HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting ,'' in 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 788--797.
  20. [20]C.-Y. Lu, Y.-S. Hung, W.-L. Chi, H.-P. Wang, C. L.-T. Tsai, Y.-C. Chang, Y.-L. Liu, T. Do, and C.-T. Lin, ``Df3dv-1k: A large-scale dataset and benchmark for distractor-free novel view synthesis,'' 2026.
  21. [21]A. Markin, V. Pryadilshchikov, A. Komarichev, R. Rakhimov, P. Wonka, and E. Burnaev, ``T-3dgs: Removing transient objects for 3d scene reconstruction,'' 2024.
  22. [22]J. Kulhanek, S. Peng, Z. Kukelova, M. Pollefeys, and T. Sattler, `` WildGaussians: 3D Gaussian Splatting In the Wild ,'' in Advances in Neural Information Processing Systems 37, 2024, pp. 21\,271--21\,288.
  23. [23]Y. Wang, M. Klasson, M. Turkulainen, S. Wang, J. Kannala, and A. Solin, `` DeSplat: Decomposed Gaussian Splatting for Distractor-Free Rendering ,'' in 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 722--732.
  24. [24]R. Wang, Q. Lohmeyer, M. Meboldt, and S. Tang, `` DeGauss: Dynamic-Static Decomposition with Gaussian Splatting for Distractor-Free 3D Reconstruction ,'' in 2025 IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 6294--6303.
  25. [25]H. Ling, X. Xu, Y. Sun, and Q. Sun, `` OCSplats: Observation Completeness Quantification and Label Noise Separation in 3DGS ,'' in 2025 IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 25\,680--25\,689.
  26. [26]C. LI, Z. Shi, Y. Lu, W. He, and X. Xu, `` Robust Neural Rendering in the Wild with Asymmetric Dual 3D Gaussian Splatting ,'' in Advances in Neural Information Processing Systems 38, 2025, pp. 55\,882--55\,907.
  27. [27]P. Asthana, A. Hanson, A. Tu, T. Goldstein, M. Zwicker, and A. 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, 2026.
  28. [28]Y. Yan, H. Lin, C. Zhou, W. Wang, H. Sun, K. Zhan, X. Lang, X. Zhou, and S. Peng, `` Street Gaussians: Modeling Dynamic Urban Scenes with Gaussian Splatting ,'' in Computer Vision -- ECCV 2024, 2024, pp. 156--173.
  29. [29]X. Zhou, Z. Lin, X. Shan, Y. Wang, D. Sun, and M.-H. Yang, `` DrivingGaussian: Composite Gaussian Splatting for Surrounding Dynamic Autonomous Driving Scenes ,'' in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 21\,634--21\,643.
  30. [30]J. Yang, B. Ivanovic, O. Litany, X. Weng, S. W. Kim, B. Li, T. Che, D. Xu, S. Fidler, M. Pavone, and Y. Wang, `` EmerNeRF: Emergent Spatial-Temporal Scene Decomposition via Self-Supervision ,'' in The Twelfth International Conference on Learning Representations, 2024.
  31. [31]A. Tonderski, C. Lindström, G. Hess, W. Ljungbergh, L. Svensson, and C. Petersson, `` NeuRAD: Neural Rendering for Autonomous Driving ,'' in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 14\,895--14\,904.
  32. [32]J. L. Schonberger and J.-M. Frahm, `` Structure-from-Motion Revisited ,'' in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 4104--4113.
  33. [33]F. Yu, H. Chen, X. Wang, W. Xian, Y. Chen, F. Liu, V. Madhavan, and T. Darrell, `` BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning ,'' in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 2636--2645.

Paper details

vlnc:W_4dc229dj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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