betaFeedback

WorksComputer Science

GaussianMamba: Geometric and Relational State Space Scanning of 3D Gaussian Primitives

Abstract

3D semantic occupancy prediction requires representing complex spatial structures and modeling contextual relationships across a scene. 3D Gaussian primitives provide a sparse and adaptive representation of scene geometry and semantics, while Mamba offers efficient sequence modeling for capturing dependencies among primitives. However, Gaussian primitives are irregular and unordered, making sequence construction challenging, as conventional spatial ordering may overlook their anisotropic geometry and feature-semantic relationships. In this paper, we present GaussianMamba, a dual-branch Mamba framework that explicitly incorporates Gaussian-specific geometric and feature-semantic relationships into complementary scanning strategies. The first branch refines a locality-preserving traversal using anisotropic Gaussian support derived from primitive scale and orientation. The second branch reorders spatially organized groups according to feature-semantic affinity and adjusts local features using geometry-weighted semantic differences. The resulting sequences are processed by separate Mamba modules, and their features are fused through a support-aware mechanism that incorporates local feature discrepancies arising from the distinct scanning paths. Experiments on nuScenes-SurroundOcc and SSCBench-KITTI-360 demonstrate consistent improvements over GaussianFormer and GaussianFormer-2 in both IoU and mIoU.

Introduction

3D semantic occupancy prediction jointly reasons about geometry and semantics to produce a dense, scene-level description of foreground objects and background structures, making it useful for autonomous driving [26, 27]. Existing methods differ in their spatial representation: BEV is efficient but compresses vertical information, voxels preserve explicit 3D structure at higher cost, and TPV retains spatial information through orthogonal planar projections [10, 26].

Recent methods instead represent scenes with sparse 3D Gaussian primitives that carry spatial and semantic attributes [11, 12]. Regardless of the representation, context exchange among scene elements is typically local: Gaussian refinement relies on sparse convolution and deformable attention over nearby primitives [11], whereas dense interaction through self-attention scales quadratically with the number of elements [13].

Selective state-space models such as Mamba [8] provide long-range context at linear cost, but applying them to spatial data requires serializing the elements into scan paths. Prior work traverses image features along multiple directions [19], reorders voxel or planar occupancy features with space-filling or structured traversals [5, 13], and orders points along space-filling curves or coordinate axes [17, 28, 29]. These strategies primarily rely on spatial layouts, point coordinates, or feature-adaptive operations, without exploiting the anisotropic support of Gaussian primitives.

Extending these strategies to 3D Gaussians is less direct. Gaussian primitives form an irregular, unordered set, each defined by a position, anisotropic scale, orientation, and opacity [11, 12], so the representation provides no predefined traversal order. Retaining the original primitive indices need not preserve spatial relationships, while coordinate-based sorting may introduce discontinuities and overlook the directional support of anisotropic primitives. Moreover, spatial proximity alone does not capture relationships between learned Gaussian features or their semantic predictions. Since a scan order defines sequential neighborhoods, different orderings yield different adjacency patterns for the same Gaussian set (Figure 1). Existing Gaussian-based Mamba modeling [7] does not explicitly address support-aware or feature-semantic scan construction. As geometric support and feature-semantic affinity may favor different neighborhoods, a single traversal may not effectively preserve both; this motivates separate scans for the anisotropic geometry and feature-semantic relationships of Gaussian primitives, together with a mechanism to integrate them.

Figure 1. Comparison of Gaussian ordering strategies. Different traversal paths establish different sequential neighborhoods among the same primitives.

To address these challenges, we introduce GaussianMamba, which incorporates geometry- and relation-aware scanning into the iterative Gaussian refinement architecture for camera-only 3D semantic occupancy prediction. The framework constructs two sequences: one emphasizes spatial continuity and anisotropic Gaussian support, while the other combines feature-semantic group routing with support-weighted boundary correction. Each sequence is processed by its own Mamba branch, and the branch outputs are integrated by a support-aware fusion mechanism that compares how they vary along their respective scan paths. The main contributions are summarized as follows:

  • We introduce Contact-Guided Spatial Sequencing (CGSS), which incorporates anisotropic Gaussian support into sequence construction through constrained local reordering that accounts for primitive scale and orientation while maintaining spatial continuity.

  • We design Directional Relation-Aware Semantic Scanning (DRSS), which combines feature-semantic group routing with support-weighted boundary correction to capture relationships beyond geometric adjacency.

  • We propose Gaussian Support-Discrepancy Fusion (GSDF), which integrates the two scan representations through the support-weighted discrepancy between their scan-local feature variations.

Related Work

3D Gaussian Representations for 3D Semantic Occupancy Prediction

Recent occupancy methods adopt 3D Gaussian primitives as sparse, adaptive alternatives to dense voxel grids. GaussianFormer [11] represents the scene with semantic Gaussians that are iteratively refined from image features, and GaussianFormer-2 [12] interprets each Gaussian as a local occupancy distribution whose contributions are combined through probabilistic superposition. For multimodal perception, GaussianFormer3D [31] incorporates LiDAR priors and 3D deformable attention, Gau-Occ [21] uses LiDAR-initialized anchors and geometry-aligned fusion, and GaussianOcc3D [7] combines LiDAR-aware aggregation with adaptive cross-modal fusion.

Other work exploits relationships among Gaussian primitives. ODG [22] separates static and dynamic components with dual Gaussian query sets, GraphGSOcc [23] models inter-Gaussian dependencies through geometric and semantic graphs whose neighborhoods adapt to Gaussian size and semantic similarity, and GaussianWorld [32] propagates Gaussians over time for streaming occupancy. These methods emphasize hierarchical refinement, graph connectivity, or temporal propagation; we instead organize inter-Gaussian relationships as scan orders for selective state-space processing, without constructing explicit graph interactions.

Mamba-based Occupancy and Spatial Serialization

Existing Mamba-based occupancy methods differ mainly in how spatial features are serialized. OccMamba [13] serializes voxel features with a height-prioritized 2D Hilbert expansion and applies hierarchical Mamba blocks for multi-scale context. MambaOcc [24] operates on compact BEV features and addresses token-order sensitivity through local adaptive reordering with deformable sampling. FMOcc [4] applies Plane Selective SSMs to TPV features to suppress redundant empty-space information. GaussianOcc3D [7] applies a Gauss-Mamba head to Gaussian primitives for state-space context modeling, without explicitly addressing support-aware or feature-semantic scan construction.

Serialization has also been studied for unordered point sets: PointMamba [17] and Point Transformer V3 [28] order points along space-filling (Hilbert, Z-order) curves, and Point Cloud Mamba [29] uses a consistent traverse serialization with axis-permuted variants. Across grid- and point-based methods, scan orders are primarily derived from predefined spatial traversals or feature-space adaptation, whereas Gaussian primitives explicitly parameterize an anisotropic extent and orientation. We employ a shared support-normalized separation measure to guide geometric sequencing, weight feature-semantic boundary correction, and drive discrepancy-based fusion.

Method

Framework Overview

Our framework builds upon the Gaussian representation and iterative refinement architecture of GaussianFormer [11], incorporating geometry- and relation-aware sequence modeling for camera-only 3D semantic occupancy prediction (Figure 2). Input images are processed by a dataset-specific CNN backbone and a multi-scale feature neck, whose features guide iterative Gaussian refinement.

Figure 2. Overview of the proposed GaussianMamba framework. Multi-view image features guide iterative Gaussian refinement. CGSS and DRSS construct geometry- and relation-aware scanning sequences, whose Mamba outputs are integrated through Gaussian Support-Discrepancy Fusion (GSDF). The refined Gaussians are splatted into a 3D semantic occupancy grid.

We represent a scene using NN Gaussian primitives {Gi}i=1N\{\mathcal{G}_i\}_{i=1}^{N}, each characterized by its center μi∈R3\boldsymbol{\mu}_i \in\mathbb{R}^{3}, anisotropic scale si∈R>03\mathbf{s}_i \in\mathbb{R}^{3}_{>0}, rotation Ri∈SO(3)\mathbf{R}_i \in\mathrm{SO}(3), stored as a unit quaternion (covariance Σi=Ridiag⁡(si2)RiT\mathbf{\Sigma}_i = \mathbf{R}_i\operatorname{diag}(\mathbf{s}_i^{2})\mathbf{R}_i^{\mathsf{T}}), opacity αi∈[0,1]\alpha_i \in[0,1], semantic probability vector pi∈RS\mathbf{p}_i \in\mathbb{R}^{S} (the softmax of its current semantic logits), and instance feature fi∈RD\mathbf{f}_i \in\mathbb{R}^{D}; the features form F∈RN×D\mathbf{F} \in\mathbb{R}^{N \times D}.

A state-space model processes its input as a sequence, so the scanning order determines which primitives become neighbors and exchange context most directly. We therefore adapt the scanning layer to Gaussian primitives at three points. Contact-Guided Spatial Sequencing (CGSS) adapts the scanning order using geometry. Directional Relation-Aware Semantic Scanning (DRSS) adapts the order using features and semantics, and corrects the features at predicted semantic boundaries. Gaussian Support-Discrepancy Fusion (GSDF) adapts how the two scans are fused. All three use the same geometric quantity, the support-normalized separation between two Gaussians (Eq. (2)).

For each branch b∈{g,r}b \in\{g,r\} (CGSS and DRSS, respectively), let Pb\mathcal{P}_b apply the scanning permutation πb\pi_b of that branch. The branch output is

Yb=Pb−1[Mb(Pb(LN⁡(F)+Δb))],(1)\mathbf{Y}_b = \mathcal{P}_b^{-1}\left[\mathcal{M}_b\left(\mathcal{P}_b\left(\operatorname{LN}(\mathbf{F})+\Delta_b\right)\right)\right], \tag*{(1)}

where LN⁡\operatorname{LN} denotes layer normalization, Mb\mathcal{M}_b an independently parameterized unidirectional Mamba [8] module, and Δb∈RN×D\Delta_b \in\mathbb{R}^{N \times D} a branch-specific feature correction: CGSS uses none (Δg=0\Delta_g = 0), while Δr\Delta_r introduces support-weighted semantic boundary information. The inverse permutation restores the original Gaussian indices; the permutations themselves receive no gradient. Since the two branches traverse the primitives in different orders, each primitive receives context from two generally different sets of predecessors.

After each refinement stage, a scanning layer consisting of the two Mamba branches and the GSDF module is applied. The refined Gaussians are finally converted into voxel-wise semantic occupancy predictions through Gaussian-to-voxel splatting.

Contact-Guided Spatial Sequencing

Gaussian primitives have no inherent ordering. A spatial traversal based on Gaussian centers provides an initial organization, but center proximity alone does not fully characterize the relationships between anisotropic primitives. In particular, at the same center distance, two Gaussians may overlap or be clearly separated, depending on their scales and orientations. To account for these properties, we introduce Contact-Guided Spatial Sequencing (CGSS, Figure 3), which constructs Gaussian-specific geometric adjacency while maintaining the broader spatial organization of the sequence.

Figure 3. Contact-Guided Spatial Sequencing (CGSS). Local reordering incorporates anisotropic Gaussian support while limiting deviations from the initial spatial traversal, without enforcing semantic grouping.

We first quantize the Gaussian centers and construct an initial traversal using a 3D Hilbert space-filling curve, whose spatial locality provides a suitable initialization for neighborhood-based refinement. Although directional Gaussian support (defined below) offers an additional criterion for determining adjacency, optimizing this measure alone does not ensure spatial continuity across the sequence. CGSS therefore retains the initial traversal as a spatial reference and introduces constrained local refinement.

For two Gaussians GiG_i and GjG_j, we define their center distance as dij=∥μj−μi∥2d_{ij} = \lVert\boldsymbol{\mu}_j - \boldsymbol{\mu}_i \rVert_2 and their normalized displacement direction as nij=(μj−μi)/(dij+ϵ)\boldsymbol{n}_{ij} = (\boldsymbol{\mu}_j - \boldsymbol{\mu}_i)/(d_{ij} + \epsilon), with a small constant ϵ\epsilon. With the rotation matrix RiR_i, the directional support of Gaussian ii and the support-normalized separation between the two primitives are computed as

ρi(nij)=∥diag⁡(si)RiTnij∥2,gij=dijρi(nij)+ρj(nij)+ϵ.(2)\rho_i(\boldsymbol{n}_{ij}) = \left\lVert\operatorname{diag}(\boldsymbol{s}_i)R_i^{\mathsf{T}}\boldsymbol{n}_{ij} \right\rVert_2, \qquad g_{ij} = \frac{d_{ij}}{\rho_i(\boldsymbol{n}_{ij}) + \rho_j(\boldsymbol{n}_{ij}) + \epsilon}. \tag*{(2)}

The scalar ρi\rho_i is the standard deviation of Gaussian ii projected onto nij\boldsymbol{n}_{ij}, i.e., its one-standard-deviation extent along that direction. Consequently, gij∈R≥0g_{ij} \in\mathbb{R}_{\geq0} measures center separation relative to the combined directional supports of the two primitives: gij≤1g_{ij} \leq1 means that the one-standard-deviation extents of the two Gaussians overlap when projected onto the line connecting their centers, which we use as a proxy for contact, whereas gij>1g_{ij} > 1 implies that their one-standard-deviation ellipsoids are disjoint. It enters the three components through different monotone mappings (Eqs. (3), (5), and (7)).

To incorporate this measure while limiting deviations from the initial spatial traversal, we define an adjacency cost:

cij=log⁡(1+gij)+dijd0+∣hi−hj∣W,(3)c_{ij} = \log(1 + g_{ij}) + \frac{d_{ij}}{d_0} + \frac{\lvert h_i - h_j\rvert}{W}, \tag*{(3)}

where cij≥0c_{ij} \geq0 is the adjacency cost, hih_i and hjh_j denote the initial sequence ranks, WW is the local refinement window size, and d0d_0 is a spatial normalization factor. The first term favors pairs with a small support-normalized separation, while the normalized distance and rank terms penalize large spatial transitions and adjacency between primitives that were distant in the initial traversal.

The sequence is refined through adjacent swaps within non-overlapping local windows of WW primitives, whose endpoints are kept fixed, so no primitive leaves its window. Each candidate swap is evaluated by comparing the sum of adjacency costs over the affected neighboring pairs. A swap is accepted only if it decreases this cost while satisfying constraints on both the total and maximum affected edge lengths, which restrict changes in spatial connectivity. After a fixed number of local sweeps, the resulting permutation πg\pi_g determines the feature sequence processed by the CGSS Mamba module.

Directional Relation-Aware Semantic Scanning

CGSS establishes Gaussian adjacency through directional support, but this criterion does not explicitly capture relationships between learned instance features or predicted semantic distributions. Moreover, Gaussian feature relationships can be considered at different levels of granularity: group-level associations describe relationships across spatial regions, whereas interactions between individual primitives characterize local semantic differences. To incorporate these complementary relationships, we introduce Directional Relation-Aware Semantic Scanning (DRSS, Figure 4), which combines relation-aware group routing with support-weighted semantic boundary correction while retaining a structured spatial traversal.

Figure 4. Directional Relation-Aware Semantic Scanning (DRSS). Feature–semantic affinity guides local group routing, while support-weighted boundary correction refines Gaussian features.

Relation-aware group routing. We first construct a directional (XYZ) spatial traversal from the quantized Gaussian centers. The traversal prioritizes the xx-axis, alternating the yy direction between consecutive xx slices and the zz direction between traversed yy rows; sorting the primitives along this traversal establishes the initial serpentine sequence. This spatial organization provides a consistent basis for grouping primitives and avoids an unrestricted global ordering driven solely by semantic predictions.

The ordered primitives are partitioned into groups of SS consecutive Gaussians, with KK groups forming a macro-block. Each group uu is represented by an ℓ2\ell_2-normalized mean instance feature fˉu∈RD\bar{\boldsymbol{f}}_u \in\mathbb{R}^{D}, an ℓ2\ell_2-normalized semantic descriptor pˉu∈RDs\bar{\mathbf{p}}_u \in\mathbb{R}^{D_s}, and a mean spatial center μˉu∈R3\bar{\boldsymbol{\mu}}_u \in\mathbb{R}^{3}. The semantic descriptor aggregates predicted class probabilities from confident Gaussians only (maximum class probability ≥τc\ge\tau_c), reducing the contribution of uncertain predictions to group relationships.

To reorganize groups according to both their represented content and spatial structure, we define the pairwise affinity

Auv=fˉuTfˉv+muvpˉuTpˉv−min⁡(∥μˉu−μˉv∥2da,3)−∣u−v∣K−1,(4)A_{uv} = \bar{\mathbf{f}}_u^{\mathsf{T}}\bar{\mathbf{f}}_v + m_{uv}\bar{\mathbf{p}}_u^{\mathsf{T}}\bar{\mathbf{p}}_v - \min\left(\frac{\lVert\bar{\boldsymbol{\mu}}_u-\bar{\boldsymbol{\mu}}_v\rVert_2}{d_a},3\right) - \frac{\lvert u-v\rvert}{K-1}, \tag*{(4)}

where Auv∈RA_{uv} \in\mathbb{R} measures the affinity between groups uu and vv of the same macro-block, muv∈{0,1}m_{uv} \in\{0,1\} is a semantic reliability gate, and dad_a is a distance normalization factor. The descriptor similarities capture feature and semantic agreement, while the bounded spatial-distance term and normalized rank difference keep the routing spatially conservative. The semantic term is disabled (muv=0m_{uv}=0) when either group has insufficient confident predictions or when the scene-level predicted class distribution is overwhelmingly concentrated in a single class.

Within each macro-block, routing begins from the first group and greedily selects the highest-affinity unvisited group until all interior groups have been assigned. The final group remains fixed, preserving the macro-block endpoints. Group-internal Gaussian ordering and the sequence of macro-blocks are also retained; primitives outside complete macro-blocks keep their initial serpentine order.

Support-weighted boundary correction. While group-level routing models relationships through aggregated descriptors, such descriptors do not explicitly encode semantic differences between individual neighboring primitives. To complement this group-level organization, we introduce a support-weighted semantic boundary correction that modifies individual Gaussian features rather than their ordering.

Candidate neighbors are obtained from the Hilbert traversal used to initialize CGSS, which serves here only to find spatially close pairs. A fixed number of preceding and following candidates are examined, retaining only pairs satisfying spatial distance and semantic confidence constraints.

For an eligible pair (i,j)(i,j), semantic disagreement is measured using the total-variation distance tij=12∥pi−pj∥1t_{ij}=\frac{1}{2}\lVert\mathbf{p}_i-\mathbf{p}_j\rVert_1, where tij∈[0,1]t_{ij}\in[0,1]. To incorporate the geometric relationship between these primitives, we combine semantic disagreement with spatial proximity, opacity, and the support-normalized separation gijg_{ij} of Eq. (2):

ηij=αiαjexp⁡[−12(max⁡(gij−1,0))2],\eta_{ij}=\sqrt{\alpha_i\alpha_j}\exp\left[-\frac{1}{2}\left(\max(g_{ij}-1,0)\right)^2\right],
wij=tijexp⁡[−(dijr0)2]η^ij.(5)w_{ij}=t_{ij}\exp\left[-\left(\frac{d_{ij}}{r_0}\right)^2\right]\hat{\eta}_{ij}. \tag*{(5)}

Here, ηij∈[0,1]\eta_{ij}\in[0,1] denotes the Gaussian contact factor, which decays as the support-normalized separation exceeds one, and r0r_0 is the boundary neighborhood radius. The normalized contact factor η^ij\hat{\eta}_{ij} is ηij\eta_{ij} divided by its weighted scene-level average and clipped. The resulting score wij≥0w_{ij}\ge0 is therefore large when two nearby Gaussians in contact are predicted to belong to different classes, i.e., at a likely semantic boundary. The score is treated as a constant; gradients flow only through the feature difference in Eq. (6).

For each Gaussian ii, we select the eligible neighbor j∗(i)j^*(i) with the highest boundary score and compute

Δr,i=wi,j∗(i)[LN⁡(fi)−LN⁡(fj∗(i))],(6)\Delta_{r,i}=w_{i,j^*(i)}\left[\operatorname{LN}(\mathbf{f}_i)-\operatorname{LN}(\mathbf{f}_{j^*(i)})\right], \tag*{(6)}

where Δr,i∈RD\Delta_{r,i}\in\mathbb{R}^{D} is the boundary-sensitive feature correction for primitive ii. That is, the scan input of ii is moved away from the normalized feature of its neighbor across the predicted boundary, in proportion to the boundary score. The correction is zero if no eligible neighbor is found, and it is added to the normalized Gaussian features before the relation-aware permutation πr\pi_r is applied Eq. (1).

Gaussian Support-Discrepancy Fusion

CGSS and DRSS establish different neighborhood relationships among the same Gaussian primitives through geometry- and relation-aware scanning, respectively. Concatenation integrates the two outputs of each primitive but does not explicitly compare how each output varies among its sequence neighbors. To account for these differences during feature integration, we introduce Gaussian Support-Discrepancy Fusion (GSDF), which incorporates scan-dependent feature discrepancies and anisotropic Gaussian support.

Let Yg\mathbf{Y}_g and Yr\mathbf{Y}_r denote the outputs of CGSS and DRSS, respectively, after restoring the original Gaussian indexing. For branch bb, b′b' denotes the other branch. For each Gaussian ii, Nb(i)\mathcal{N}_b(i) contains its immediately preceding and following primitives in the scanning order of branch bb, when available. Using the support-normalized separation gijg_{ij} defined in Eq. (2), we define the Gaussian support affinity as kij=exp⁡(−gij)k_{ij} = \exp(-g_{ij}) and compute

ξb,i=∑j∈Nb(i)kij(Yb,j−Yb,i)1+∑j∈Nb(i)kij,qb,i=1∣Nb(i)∣∑j∈Nb(i)kij.(7)\xi_{b,i} = \frac{\sum_{j \in\mathcal{N}_b(i)} k_{ij}(\mathbf{Y}_{b,j} - \mathbf{Y}_{b,i})}{1 + \sum_{j \in\mathcal{N}_b(i)} k_{ij}}, \qquad q_{b,i} = \frac{1}{\lvert\mathcal{N}_b(i) \rvert} \sum_{j \in\mathcal{N}_b(i)} k_{ij}. \tag*{(7)}

Here, ξb,i∈RD\xi_{b,i} \in\mathbb{R}^{D} captures support-weighted local feature variation along each scanning path, while qb,iq_{b,i} down-weights the adjustment in Eq. (8) where the scan neighbors are far apart relative to their supports.

The difference between the branch-specific feature contrasts is incorporated into the Gaussian representations before their integration:

Y~b,i=Yb,i+λqb,i(ξb,i−ξb′,i),fiout=Wf[Y~g,i;Y~r,i]+bf.(8)\widetilde{\mathbf{Y}}_{b,i} = \mathbf{Y}_{b,i} + \lambda q_{b,i}(\xi_{b,i} - \xi_{b',i}), \qquad \mathbf{f}^{\mathrm{out}}_i = \mathbf{W}_f[\widetilde{\mathbf{Y}}_{g,i};\widetilde{\mathbf{Y}}_{r,i}] + \mathbf{b}_f. \tag*{(8)}

Here, Wf∈RD×2D\mathbf{W}_f \in\mathbb{R}^{D \times2D} and bf∈RD\mathbf{b}_f \in\mathbb{R}^{D} denote the learned projection parameters, and λ=tanh⁡(θ)\lambda= \tanh(\theta) is a learnable scalar initialized to zero, so that GSDF starts from plain concatenation fusion. Equivalently, the difference between the two adjusted branches changes by λ(qg,i+qr,i)(ξg,i−ξr,i)\lambda(q_{g,i} + q_{r,i})(\xi_{g,i} - \xi_{r,i}): GSDF amplifies or suppresses the inter-branch difference where the two scans see different local variation. The fused feature fiout\mathbf{f}^{\mathrm{out}}_i replaces fi\mathbf{f}_i as the input to the next block.

Experiment

Dataset and Metric

The nuScenes-SurroundOcc dataset [2] contains 1K surround-view driving scenes collected in Boston and Singapore, officially split into 700 training, 150 validation, and 150 testing scenes. Each scene spans 20 seconds with annotations at 2 Hz and includes camera, LiDAR, radar, and IMU data. We adopt the 3D semantic occupancy annotations from SurroundOcc [27] for supervision and evaluation. The annotated volume covers [−50,50][-50,50] m along the x- and y-axes and [−5,3][-5,3] m along the z-axis, with a voxel size of 0.5 m. The 18 occupancy labels comprise 16 semantic categories, one empty class, and one noise class.

The KITTI-360 dataset [18] provides over 320K images of suburban driving environments, captured using multiple cameras and LiDAR sensors. We use only the ego vehicle’s left-camera images as input. For occupancy supervision and evaluation, we adopt SSCBench-KITTI-360 [15], officially split into 7 training, 1 validation, and 1 testing sequence, containing 8,487, 1,812, and 2,566 keyframes, respectively. The annotated region spans 51.2×51.2×6.451.2 \times51.2 \times6.4 m3^3 ahead of the vehicle, discretized into 256×256×32256 \times256 \times32 voxels with 19 labels, including 18 semantic classes and one empty class.

The evaluation metrics adhere to common practice, namely mean Intersection-over-Union (mIoU) and Intersection-over-Union (IoU):

mIoU⁡=1∣C′∣∑i∈C′TPiTPi+FPi+FNi,(9)\operatorname{mIoU} = \frac{1}{\lvert\mathcal{C}' \rvert} \sum_{i \in\mathcal{C}'} \frac{TP_i}{TP_i + FP_i + FN_i}, \tag*{(9)}
IoU⁡=TP≠c0TP≠c0+FP≠c0+FN≠c0,(10)\operatorname{IoU} = \frac{TP_{\ne c_0}}{TP_{\ne c_0} + FP_{\ne c_0} + FN_{\ne c_0}}, \tag*{(10)}

where C′\mathcal{C}' is the set of non-empty classes, c0c_0 is the empty class, and TP, FP, and FN count the true-positive, false-positive, and false-negative voxels.

Implementation Details

For nuScenes-SurroundOcc [2] and KITTI-360 [18], we use input image resolutions of 900×1600900 \times1600 and 376×1408376 \times1408. Following GaussianFormer [11], we initialize the image backbones using the same pretrained checkpoints: ResNet101-DCN [9] with FCOS3D pretrained weights [25] for nuScenes-SurroundOcc, and ImageNet-pretrained [6] ResNet50 [9] for KITTI-360. On nuScenes-SurroundOcc, GaussianFormer uses 144K Gaussian primitives, whereas GaussianFormer-2 employs 12.8K primitives with 128-dimensional features. For KITTI-360, we use 38.4K Gaussians. The scanning layers use W=32W=32, d0=3d_0=3 m, S=128S=128, K=8K=8, τc=0.16\tau_c=0.16, da=10d_a=10 m, and r0=2r_0=2 m (two sweeps; total/maximum edge growth ≤5%/10%\leq5\%/10\%; gate: ≥S/8\geq S/8 confident Gaussians per group, dominant class ≤90%\leq90\%; six candidates per side). We train with the unchanged cross-entropy and Lovász-softmax [1] losses of GaussianFormer and GaussianFormer-2 [11, 12] using AdamW [20] with a weight decay of 0.01 and a peak learning rate of 2×10−42 \times10^{-4}, following a cosine annealing schedule. Training is conducted for 20 epochs with a batch size of 8 on nuScenesSurroundOcc and 30 epochs with a batch size of 4 on KITTI-360.

Results

#B and #G denote Mamba branches and Gaussian counts; † indicates three-seed accuracy averages. All ablations use 144K Gaussians. Dual branches use independent Mamba parameters, defaulting to Hilbert scanning and concatenation fusion unless specified otherwise. Latency is measured on an NVIDIA RTX 6000 Ada and reported per scanning layer, which is applied after each of the four refinement stages; ordering ablations cover scanning and Mamba processing, while fusion parameters and latency are module-specific.

Results on nuScenes-SurroundOcc. Table 1 shows that GaussianMamba improves GaussianFormer by 2.58 IoU and 1.45 mIoU points, surpassing SurroundOcc, while GaussianFormer-2 gains 2.42 and 1.10 points, achieving the highest IoU and mIoU in the comparison. Improvements span all 16 classes with GaussianFormer-2 and 13 with GaussianFormer. Notably, drivable surface gains 6.25 points with GaussianFormer and 3.14 with GaussianFormer-2, with consistent improvements in terrain, suggesting benefits for spatially extended regions.

MethodIoUmIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. suf.other flatsidewalkterrainmanmadevegetation
BEVFormer [16]30.5016.7514.226.5823.4628.288.6610.776.644.0511.2017.7837.2818.0022.8822.1713.8022.21
TPVFormer [10]11.5111.6616.147.1722.6317.138.8311.3910.468.239.4317.028.0713.6413.8510.344.907.37
TPVFormer* [10]30.8617.1015.965.3123.8627.329.798.747.095.2010.9719.2238.8721.2524.2623.1511.7320.81
OccFormer [30]31.3919.0318.6510.4123.9230.2910.3114.1913.5910.1312.4920.7738.7819.7924.1922.2113.4821.35
SurroundOcc [27]31.4920.3020.5911.6828.0630.8610.7015.1414.0912.0614.3822.2637.2923.7024.4922.7714.8921.86
GaussianFormer [11]29.8319.1019.5211.2626.1129.7810.4713.8312.588.6712.7421.5739.6323.2824.4622.999.5919.12
GaussianFormer + Ours†32.4120.5519.3614.0426.3232.108.4615.9613.349.2712.5722.0145.8826.0427.7225.7210.1819.76
GaussianFormer-2 [12]30.5620.0220.1512.9927.6130.2311.1915.3112.649.6313.3122.2639.6823.4725.6223.2012.2520.73
GaussianFormer-2 + Ours†32.9821.1220.2714.2529.4531.6411.9216.6513.0410.4513.9523.0342.8223.5026.8225.5912.7521.86

Table 1. Surround view 3D semantic occupancy prediction results on nuScenes-SurroundOcc. TPVFormer* is supervised by dense occupancy annotations.

Results on SSCBench-KITTI-360. For monocular occupancy prediction (Table 2), GaussianMamba improves GaussianFormer by 2.49 IoU and 1.03 mIoU points, and GaussianFormer-2 by 2.20 and 0.90 points, the latter achieving the highest overall scores. Gains are particularly evident in spatially extended classes, including road, sidewalk, building, and vegetation (+2.2 to +3.8 points), while individual class regressions remain within 0.35 points. These results suggest consistent benefits across different Gaussian representations and benchmarks.

MethodInputIoUmIoUcarbicyclemotorcycletruckother-veh.personroadparkingsidewalkother-grndbuildingfencevegetationterrainpoletraf.-signother-struct.other-object
MonoScene [3]C37.8712.3119.340.430.588.022.030.8648.3511.3828.133.2232.893.5326.1516.756.925.674.203.09
Voxformer [14]C38.7611.9117.841.160.894.562.061.6347.019.6727.212.8931.184.9728.9914.696.516.923.792.43
TPVFormer [10]C40.2213.6421.561.091.378.062.572.3852.9911.9931.073.7834.834.8030.0817.517.465.865.482.70
OccFormer [30]C40.2713.8122.580.660.269.893.822.7754.3013.4431.533.5536.424.8031.0019.517.778.516.954.60
GaussianFormer [11]C35.3812.9218.931.024.6218.077.593.3545.4710.8925.035.3228.445.6829.548.622.992.329.515.14
GaussianFormer + Ours†C37.8713.9518.811.325.0220.177.343.8049.2912.2027.875.1431.276.3231.958.303.392.8211.065.03
GaussianFormer-2 [12]C38.3713.9021.082.554.2112.415.731.5954.1211.0432.313.3432.014.9828.9417.333.575.485.883.54
GaussianFormer-2 + Ours†C40.5714.8023.082.794.0713.315.521.8556.3612.6035.383.1634.685.4332.2616.984.125.235.703.88

Table 2. Monocular 3D semantic occupancy prediction results on SSCBench-KITTI-360.

Branch capacity vs. scan order. Table 3 and Table 4 distinguish the effects of branch capacity and Gaussian sequencing. Duplicating the Hilbert scan yields only 0.05 mIoU improvement, and replacing the duplicate with a reversed (bidirectional) or an XYZ scan adds only 0.05 and 0.19 more. With identical dual-branch capacity, CGSS and DRSS outperform Hilbert + XYZ by 0.83 mIoU at 7.6 ms additional latency. Their joint improvement (1.02) also exceeds the sum of their individual gains (0.25 and 0.39), suggesting complementary geometric and semantic context.

Configuration#BCGSSDRSSGSDFIoU ↑mIoU ↑
GaussianFormer [11]029.8319.10
Single Mamba130.0519.18
Dual Mamba†230.2419.23
+ GSDF2✓30.2719.25
+ CGSS2✓30.8819.48
+ DRSS2✓30.6819.62
+ CGSS + DRSS†2✓✓31.8220.25
+ CGSS + GSDF2✓✓31.0319.59
+ DRSS + GSDF2✓✓30.9119.72
GaussianMamba (Ours)†2✓✓✓32.4120.55

Table 3. Component ablation on nuScenes-SurroundOcc.

Scanning Strategy#BIoU ↑mIoU ↑Latency (ms) ↓
XYZ129.9619.066.99
Morton130.0019.136.10
Hilbert130.0519.1810.68
Hilbert + Hilbert230.2419.2316.16
Hilbert + Reversed Hilbert230.2919.2816.74
Hilbert + XYZ230.4819.4217.40
CGSS + DRSS†231.8220.2524.98

Table 4. Comparison of Gaussian scanning strategies on nuScenes-SurroundOcc. All dual-branch configurations employ independently parameterized Mamba branches with concatenation-based fusion.

Gaussian support in CGSS. Table 5 isolates the contribution of anisotropic Gaussian support. Constrained reordering without support improves mIoU by 0.29 while barely changing the share of adjacent pairs in contact (12.6% to 13.4%), whereas incorporating support raises contact to 18.7% and adds a further 0.34 mIoU. Both the constrained reordering itself and the Gaussian extent beyond center proximity therefore contribute to spatial sequencing.

Geometry-Branch OrderContact (%) ↑IoU ↑mIoU ↑
Hilbert12.6030.6819.62
CGSS w/o Gaussian support13.4031.2119.91
CGSS18.7031.8220.25

Table 5. Ablation of CGSS design choices on nuScenes-SurroundOcc. The second branch uses DRSS. Contact denotes the percentage of adjacent Gaussian pairs satisfying gij≤1g_{ij} \le1. Removing support excludes log⁡(1+gij)\log(1 + g_{ij}) from the adjacency cost.

Qualitative comparison of GaussianFormer and GaussianFormer-2 with and without GaussianMamba on nuScenes-SurroundOcc

Figure 5. Qualitative comparison of GaussianFormer and GaussianFormer-2 with and without GaussianMamba on nuScenes-SurroundOcc.

DRSS components. Table 6 shows that group routing and boundary correction independently improve mIoU by 0.42 and 0.25, respectively. Their combined gain of 0.70 is close to the sum of the individual gains (0.67), indicating largely additive contributions from semantic ordering and support-aware boundary correction.

Group routingBoundary correctionIoU ↑mIoU ↑
30.9819.55
✓31.3919.97
✓31.2019.80
✓✓31.8220.25

Table 6. Ablation of DRSS components on nuScenes-SurroundOcc. The first branch uses CGSS; the first row represents plain directional scanning.

Fusion strategy. Table 7 shows that cross-attention requires 4.7×4.7 \times the parameters of concatenation for only 0.04

Fusion StrategyParams (K)IoU ↑\uparrowmIoU ↑\uparrowLatency (ms) ↓\downarrow
Addition16.5131.5419.960.61
Concatenation32.8031.8220.250.82
Cross-attention155.3232.0120.299.13
GSDF †33.2032.4120.552.80

Table 7. Comparison of feature fusion strategies for CGSS and DRSS on nuScenes-SurroundOcc.

additional mIoU. In contrast, GSDF improves mIoU by 0.30 with only 0.40K additional parameters and 2.0 ms latency overhead. Its contribution increases from 0.02 mIoU with duplicated Hilbert scans to 0.10–0.11 with one specialized branch and 0.30 with both (Table 3). Although the smaller differences should be interpreted with caution, this trend is consistent with GSDF exploiting the difference between the two scans rather than added capacity.

Computational Efficiency. Table 8 examines the trade-off between Gaussian count, efficiency, and prediction accuracy. Latency and peak GPU memory are measured during inference using trained models. Reducing Gaussian count substantially decreases memory usage and inference latency. GaussianMamba adds only 0.8M parameters with negligible memory overhead, improving both IoU and mIoU at either scale. With GaussianFormer, the gains are larger at 144K Gaussians (+2.58 IoU, +1.45 mIoU) than at 25.6K (+0.88 IoU, +0.97 mIoU); this does not generalize across backbones, however, as GaussianFormer-2 gains 2.42 IoU and 1.10 mIoU with only 12.8K Gaussians. The additional scanning operations increase inference latency by 30.61 ms and 111.42 ms at 25.6K and 144K Gaussians, respectively; the latter is consistent with four scanning layers of roughly 27.8 ms each (Table 4 and Table 7), highlighting a computational trade-off, particularly for larger Gaussian sets.

Method#GParams (M)Mem. (GB) ↓\downarrowLat. (ms) ↓\downarrowIoU ↑\uparrowmIoU ↑\uparrow
GaussianFormer144K75.552.77354.7129.8319.10
GaussianFormer + Ours†144K76.352.79466.1232.4120.55
GaussianFormer25.6K57.092.08183.5128.7216.00
GaussianFormer + Ours†25.6K57.892.11214.1229.6016.97

Table 8. Computational efficiency and performance under different Gaussian counts on nuScenes-SurroundOcc.

Conclusion

We presented GaussianMamba, a dual-branch Mamba framework that incorporates Gaussian-specific geometry and feature-semantic relationships into sequence modeling for 3D semantic occupancy prediction. CGSS accounts for anisotropic Gaussian support during spatial sequencing, DRSS captures feature-semantic relationships beyond geometric adjacency, and GSDF fuses the two scans through support-weighted scan-local discrepancies. On nuScenes-SurroundOcc and SSCBench-KITTI-360, GaussianMamba improves GaussianFormer and GaussianFormer-2 by 2.2–2.6 IoU and 0.9–1.5 mIoU points with modest parameter and memory overhead. The main limitation is increased inference latency, reaching approximately 31% overhead at 144K Gaussians. Future work will focus on more efficient Gaussian scanning to reduce this computational cost.

LLM Usage Statement

OpenAI’s ChatGPT and Anthropic’s Claude were used to assist with manuscript preparation, including grammar and spelling correction, language refinement, identification of typographical errors, and suggestions for improving clarity and consistency across the text. The authors reviewed the suggested revisions and remain responsible for the scientific content, technical claims, experimental results, and final manuscript.

References

  1. [1]Maxim Berman, Amal Rannen Triki, and Matthew B. Blaschko. The Lovász-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4413–4421, 2018.
  2. [2]Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multi-modal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11621–11631, 2020.
  3. [3]Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3981–3991. IEEE, 2022.arxiv.org/abs/2112.00726
  4. [4]Jiangxia Chen, Tongyuan Huang, and Ke Song. Fmocc: Tpv-driven flow matching for 3d occupancy prediction with selective state space model. arXiv preprint arXiv:2507.02250, 2025.
  5. [5]Yubo Cui, Zhiheng Li, Jiaqiang Wang, and Zheng Fang. Loma: Language-assisted semantic occupancy network via triplane mamba. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2609–2617, 2025.arxiv.org/abs/2412.08388
  6. [6]Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009.
  7. [7]A Doruk and Hasan F Ates. Gaussianocc3d: A gaussian-based adaptive multi-modal 3d occupancy prediction. arXiv preprint arXiv:2601.22729, 2026.arxiv.org/abs/2601.22729
  8. [8]Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. In First Conference on Language Modeling (COLM), 2024.arxiv.org/abs/2312.00752
  9. [9]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016.arxiv.org/abs/1512.03385
  10. [10]Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9223–9232. IEEE, 2023.arxiv.org/abs/2302.07817
  11. [11]Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In European Conference on Computer Vision, pages 376–393. Springer, 2024.arxiv.org/abs/2405.17429
  12. [12]Yuanhui Huang, Amonnut Thammadatakrkoon, Wenzhao Zheng, Yunpeng Zhang, Dalong Du, and Jiwen Lu. Gaussianformer-2: Probabilistic gaussian superposition for efficient 3d occupancy prediction. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27477–27486. IEEE, 2025. 1, 2, 5, 6, 7arxiv.org/abs/2412.04384
  13. [13]Heng Li, Yuenan Hou, Xiaohan Xing, Yuexin Ma, Xiao Sun, and Yanyong Zhang. Occmamba: Semantic occupancy prediction with state space models. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11949–11959. IEEE, 2025. 1, 2arxiv.org/abs/2408.09859
  14. [14]Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9087–9098. IEEE, 2023. 7arxiv.org/abs/2302.12251
  15. [15]Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, et al. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13333–13340. IEEE, 2024. 6arxiv.org/abs/2306.09001
  16. [16]Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sim, Tong Lu, Yu Qiao, and Jifeng Dai. BEVFormer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision (ECCV), pages 1–18, 2022. 5arxiv.org/abs/2203.17270
  17. [17]Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaoqing Ye, Xiao Tan, and Xiang Bai. PointMamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2arxiv.org/abs/2402.10739
  18. [18]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, 2022. 6arxiv.org/abs/2109.13410
  19. [19]Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixian Ye, Jianbin Jiao, and Yunfan Liu. Vrmamba: Visual state space model. Advances in neural information processing systems, 37:103031–103063, 2024. 1arxiv.org/abs/2401.10166
  20. [20]Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6arxiv.org/abs/1711.05101
  21. [21]Chengxin Lv, Yihui Li, Hongyu Yang, and Yunhong Wang. Gau-occ: Geometry-completed gaussians for multi-modal 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14184–14207, 2024. 2
  22. [22]Yunxiao Shi, Yinhao Zhu, Shizhong Han, Jisoo Jeong, Amin Ansari, Hong Cai, and Fatih Porikli. ODG: Occupancy prediction using dual gaussians. In Advances in Neural Information Processing Systems (NeurIPS), 2025. 2arxiv.org/abs/2506.09417
  23. [23]Ke Song, Yunhe Wu, Chunchit Siu, Huiyuan Xiong, and Qiqi Xu. Graphgsocc: Semantic-geometric graph transformer with dynamic-static decoupling for 3d gaussian splatting-based occupancy prediction. IEEE Transactions on Circuits and Systems for Video Technology, 2026. 2arxiv.org/abs/2506.14825
  24. [24]Yonglin Tian, Songlin Bai, Zhiyao Luo, Yutong Wang, Hui Zhang, Baoqing Guo, Yisheng Lv, and Fei-Yue Wang. Mambaocc: Visual state space models for bev-based occupancy prediction with local adaptive reordering. Expert Systems with Applications, page 131305, 2026. 2
  25. [25]Tai Wang, Xingye Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 913–922. IEEE, 2021. 6arxiv.org/abs/2104.10956
  26. [26]Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Opensoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 17804–17813. IEEE, 2023. 1arxiv.org/abs/2303.03991
  27. [27]Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 21672–21683. IEEE, 2023. 1, 5, 6arxiv.org/abs/2303.09551
  28. [28]Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer V3: Simpler, faster, stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4840–4851, 2024. 1, 2arxiv.org/abs/2312.10035
  29. [29]Tao Zhang, Haobo Yuan, Lu Qi, Jiangning Zhang, Qianyu Zhou, Shunping Ji, Shuicheng Yan, and Xiangtai Li. Point cloud Mamba: Point cloud learning via state space model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10121–10130, 2025. 1, 2arxiv.org/abs/2403.00762
  30. [30]Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9399–9409. IEEE, 2023. 5, 7arxiv.org/abs/2304.05316
  31. [31]Lingjun Zhao, Sizhe Wei, James Hays, and Lu Gan. Gaussianformer3d: Multi-modal gaussian-based semantic occupancy prediction with 3d deformable attention. In IEEE International Conference on Robotics and Automation (ICRA), 2026. 2arxiv.org/abs/2505.10685
  32. [32]Sicheng Zuo, Wenzhao Zheng, Yuanhui Huang, Jie Zhou, and Jiwen Lu. Gaussianworld: Gaussian world model for streaming 3d occupancy prediction. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6772–6781. IEEE, 2025. 2arxiv.org/abs/2412.10373

Paper details

vlnc:W_7wejpyj4
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

OpenAI’s ChatGPT, Anthropic’s Claude

View statement
Looking for related papers…
Contents