Pith. sign in

REVIEW 2 major objections 5 minor 67 references

FocusGS reconstructs driving scenes from sparse views by completing geometry only where it is ambiguous, cutting Gaussians by ~74% and render time by ~34%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 15:36 UTC pith:KDEUZTPV

load-bearing objection Clean engineering win: targeted completion on a depth-gradient ambiguity manifold beats Omni-Scene on quality while cutting Gaussians ~74% and latency ~34%. the 2 major comments →

arxiv 2607.04661 v1 pith:KDEUZTPV submitted 2026-07-06 cs.CV cs.AI

Targeted Structure Completion for Sparse-View 3D Reconstruction in Autonomous Driving

Guoqing Wang , Pin Tang , Xiangxuan Ren , Liping Hou , Chao Ma This is my paper
classification cs.CV cs.AI
keywords 3D Gaussian Splattingsparse-view reconstructionautonomous drivinggeometric ambiguitytargeted structure completionego-centric camerasfeed-forward Gaussians
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Sparse multi-camera views in autonomous driving have almost no overlap, so pure pixel-based Gaussian methods leave holes at occlusions while full voxel methods waste computation on simple surfaces like roads and sky. FocusGS starts with an efficient pixel-aligned base, then builds a 3D Geometric Ambiguity Manifold from depth discontinuities to mark only the uncertain regions. A lightweight module places and optimizes extra Gaussian queries strictly inside that sparse subspace. The result matches or beats prior dual-branch quality on driving benchmarks while using far fewer primitives and rendering faster. The practical payoff is a reconstruction pipeline that stays real-time-friendly without sacrificing structural completeness where it matters.

Core claim

Structural completion for sparse-view 3D Gaussians can be fully decoupled from deterministic regions: localize a sparse 3D Geometric Ambiguity Manifold from depth-gradient boundaries, then instantiate and optimize continuous Gaussian queries only inside that manifold. This targeted strategy recovers occluded geometry at the quality of uniform volumetric methods while eliminating the redundant Gaussians those methods place on flat, well-observed surfaces.

What carries the argument

The 3D Geometric Ambiguity Manifold: a sparse topological subspace obtained by thresholding and dilating depth gradients in each view, then lifting the resulting 2D bands into short 3D segments along camera rays; all subsequent completion queries are sampled and updated exclusively inside this subspace.

Load-bearing premise

The depth-gradient map (plus a fixed dilation and thickness) is assumed to mark every geometrically incomplete region that actually hurts novel-view quality; if the predicted depth is already wrong or sensor artifacts dominate, the manifold misses or over-includes the wrong places.

What would settle it

On the same nuScenes splits, replace the depth-gradient manifold with a pure random mask of equal volume or with ground-truth high-error regions and measure whether PSNR/SSIM/PCC still improve by the reported margins while keeping the same query budget; collapse of the gains would falsify the claim that the manifold is the essential localizer.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Dual-branch pixel-plus-voxel pipelines can drop their dense volumetric branch and still recover occlusions by completing only the ambiguity manifold.
  • Total Gaussian count and rendering latency become largely independent of scene volume and scale mainly with the size of the uncertain subspace.
  • The same manifold-plus-targeted-queries pattern can be dropped onto other pixel-based feed-forward predictors to raise their geometric completeness without full re-architecture.
  • Driving reconstruction systems can allocate compute adaptively at runtime by monitoring the fraction of pixels that fall inside the current ambiguity manifold.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the manifold is recomputed every few frames from streaming depth, the method naturally extends to online dynamic scenes without storing a static dense volume.
  • Adverse-weather failures suggest a natural next step: treat lens artifacts themselves as an extra ambiguity class so the completion module can ignore or inpaint them rather than bake them into 3D.
  • The fixed query budget N_q acts as a hard efficiency dial; practitioners can trade a few dB of PSNR for still lower memory by lowering N_q once the saturation point is known.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. FocusGS addresses sparse-view 3D reconstruction for ego-centric autonomous driving by decoupling structural completion from deterministic surfaces. It first builds a pixel-aligned base Gaussian representation, then derives a 2D geometric ambiguity manifold from depth-gradient discontinuities (Eqs. 3–5), lifts it into a sparse 3D uncertainty subspace via longitudinal thickness (Eqs. 6–8), and instantiates a fixed budget of continuous Gaussian queries only inside that subspace. These queries are refined by sparse self-context aggregation, deformable cross-view attention, and an MLP decoder. On nuScenes the method reports higher PSNR/SSIM/PCC and lower LPIPS than Omni-Scene while cutting total Gaussians by ~74 % and rendering latency by ~34 %; competitive or better numbers appear on RealEstate10K, and ablations isolate the contribution of targeted localization over random completion.

Significance. If the efficiency–quality claims hold under broader deployment, FocusGS supplies a practical middle path between pure pixel-based feed-forward Gaussians (fast but incomplete under low overlap) and dual-branch voxel methods (complete but redundant). The explicit geometric-ambiguity manifold and fixed-budget query design are simple, modular, and immediately usable by other sparse-view pipelines; the systematic ablations (Tables 3–8) and the public project page further raise the work’s utility for the autonomous-driving reconstruction community.

major comments (2)
  1. The central efficiency claim rests on the premise that the depth-gradient manifold (Eqs. 3–5, τ_g = 3, 3×3 dilation) plus longitudinal thickness (Eq. 7) captures essentially all regions whose completion improves novel-view metrics. Table 5 shows solid recall (0.82) against an oracle error mask, yet the paper itself documents systematic failure under rain/lens artifacts (Fig. 7 and Limitations). A quantitative stress test—e.g., controlled depth-noise injection or evaluation on a rainy nuScenes subset—would make the robustness boundary of this load-bearing assumption explicit rather than anecdotal.
  2. Table 1 reports latency and quality gains versus Omni-Scene, and Fig. 6 visualizes Gaussian counts, but the manuscript never states the absolute Gaussian budgets (base + compensatory) used for each method under identical hardware and resolution. Without those numbers the ~74 % / ~34 % reductions cannot be independently verified or compared with other sparse-view baselines that also claim efficiency.
minor comments (5)
  1. Hyper-parameters κ_rel and κ_abs that define longitudinal thickness (Eq. 7) are never given numerical values; they should appear in the implementation details or an appendix.
  2. Fig. 2 percentages (82 % flat / 18 % ambiguity) are presented without stating how the regions were segmented or whether they are averaged over the full test set; a short caption clarification would help.
  3. In §3.3 the notation switches between N_q and “Nums” (Tables 6–7); consistent symbols would improve readability.
  4. The RealEstate10K protocol (number of input views, exact train/test split) is only sketched; a one-sentence alignment with the Omni-Scene protocol would remove ambiguity.
  5. A few typographical inconsistencies remain (e.g., “F ocusGS”, “V oxel-based”, missing spaces around ~74 %).

Circularity Check

0 steps flagged

No significant circularity; FocusGS is an empirical architecture whose efficiency-quality claims are measured against held-out novel views and external baselines, not forced by construction or self-citation.

full rationale

The paper defines a 2D geometric ambiguity manifold via depth-gradient thresholding and morphological dilation (Eqs. 3-5), lifts it to a 3D uncertainty subspace (Eqs. 6-8), and samples a fixed budget of continuous Gaussian queries strictly inside that subspace for targeted completion. All quality metrics (PSNR/SSIM/LPIPS/PCC) and efficiency numbers (~74% fewer Gaussians, ~34% lower latency) are obtained by rendering novel views and comparing against independent ground-truth images and published baselines (Omni-Scene, MVSplat, etc.). Ablations (Tables 3-8) isolate the contribution of the manifold versus random completion and quantify mask recall against an oracle error mask; none of these steps reduce a claimed prediction to a fitted input or to a self-citation uniqueness theorem. The manifold is derived from the model's own depth prediction, yet supervision remains photometric/depth losses on held-out data, so the central SOTA claim is not circular by construction. No load-bearing self-citations, ansatz smuggling, or renaming of known results appear. Residual risk is ordinary domain-shift sensitivity (rain artifacts) already acknowledged by the authors, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 2 invented entities

The central efficiency claim rests on a handful of hand-chosen thresholds and budgets that define the manifold and the query set; the geometric prior that depth discontinuities coincide with reconstruction error is an empirical domain assumption validated only by the paper’s own error-map correlation figure.

free parameters (5)
  • gradient threshold τ_g = 3
    Hard threshold on depth-gradient magnitude that decides which pixels enter the 2D ambiguity manifold; set to 3 without cross-validation reported.
  • dilation kernel size k = 3
    Morphological expansion of strict boundaries into uncertainty bands; chosen as 3 after ablation, directly controls manifold volume.
  • query budget N_q = 60k
    Fixed number of continuous Gaussian queries sampled from the 3D subspace; set to 60k as the operating point after saturation study.
  • longitudinal thickness coefficients κ_rel, κ_abs
    Control the near-far extent of each lifted ray segment; values not numerically listed but required for the 3D subspace definition.
  • number of completion blocks = 4
    Depth of the sparse-conv + deformable-attention stack; set to 4 after ablation showing degradation at 1 and 6.
axioms (3)
  • domain assumption Depth discontinuities (large finite-difference gradients) are a reliable proxy for geometric ambiguity and occlusion boundaries in ego-centric driving scenes.
    Stated in §3.2 and illustrated by the error-map overlap in Fig. 4; if false under weather or dynamic objects the manifold fails.
  • ad hoc to paper A fixed global query budget sampled uniformly from the lifted subspace is sufficient to complete all critical missing structure without dense volumetric context.
    Core design choice of the targeted completion module (§3.3); justified only by the efficiency-quality trade-off experiments.
  • domain assumption Pixel-aligned base Gaussians already achieve high fidelity on deterministic surfaces (flat roads, sky), so completion can be completely decoupled from those regions.
    Motivated by the area analysis in Fig. 2 (82% flat area with PSNR 26.5); underpins the entire selective-computation claim.
invented entities (2)
  • 3D Geometric Ambiguity Manifold (O_3d) no independent evidence
    purpose: Sparse topological subspace that localizes where structural completion is allowed to operate.
    Constructed by lifting dilated 2D depth-gradient masks into ray segments; no independent geometric definition outside this paper.
  • Targeted Structure Completion module (continuous Gaussian queries + sparse self/cross aggregation) no independent evidence
    purpose: Lightweight decoder that instantiates and optimizes Gaussians strictly inside the manifold.
    New architectural block combining sparse 3D conv, deformable attention and residual MLP; purpose-built for the paper’s efficiency goal.

pith-pipeline@v1.1.0-grok45 · 19901 in / 2932 out tokens · 24833 ms · 2026-07-11T15:36:21.487335+00:00 · methodology

0 comments
read the original abstract

Reconstructing 3D scene structures from sparse, low-overlap observations remains a fundamental challenge in autonomous driving. Recent state-of-the-art frameworks achieve promising results by incorporating voxel-based Gaussians, but incur substantial computational redundancy due to a uniform volumetric processing strategy. To bridge the gap between the efficiency of pixel-based Gaussian methods and the structural completeness of voxel-based Gaussian approaches, we propose FocusGS, a simple yet effective framework that shifts the paradigm from global densification to targeted structural completion. Our central insight is that structural completion should be decoupled from deterministic regions, with computation concentrated exclusively on areas exhibiting geometric ambiguity. Specifically, FocusGS addresses the localization challenge by deriving a 3D Geometric Ambiguity Manifold to accurately isolate localized areas prone to occlusion and high geometric uncertainty. To overcome the subsequent manifold completion challenge, we design a lightweight targeted structure completion module that selectively instantiates and optimizes continuous Gaussian queries strictly within this unstructured, sparse topological subspace. Extensive experiments demonstrate that FocusGS achieves a superior efficiency-quality trade-off, advancing state-of-the-art performance on driving-centric benchmarks while naturally reducing the total number of Gaussians by ~74% and decreasing rendering time by ~34%.

Figures

Figures reproduced from arXiv: 2607.04661 by Chao Ma, Guoqing Wang, Liping Hou, Pin Tang, Xiangxuan Ren.

Figure 1
Figure 1. Figure 1: Comparison of feed-forward 3D Gaussian Splatting methods. (a, b) Prior works rely on either pixel- [5, 8] or voxel-based [44] predictors. (c) Omni-Scene [49] combines both but suffers from redundant Gaussians. (d) Our FocusGS introduces targeted struc￾ture completion based on pixel-based features, achieving high-fidelity reconstruction with minimal Gaussian overhead. approaches offer remarkable efficiency … view at source ↗
Figure 2
Figure 2. Figure 2: Quantitative analysis of pixel-based methods across different geometric areas. completion in critical regions pro￾vides a simple and principled so￾lution. Rather than uniformly gen￾erating Gaussians across the entire 3D volume, the model should selec￾tively concentrate on critical areas. We formally define these critical ar￾eas as geometric ambiguity manifold, i.e., a localized, sparse 3D topological subsp… view at source ↗
Figure 3
Figure 3. Figure 3: The overall pipeline of FocusGS. Given sparse multi-view images, we first gen￾erate a pixel-aligned base Gaussian representation for continuous visible surfaces. To handle occlusions and depth discontinuities, we explicitly localize regions of high uncer￾tainty by constructing a geometric ambiguity manifold. A targeted structure completion module then initializes and optimizes Gaussian queries exclusively … view at source ↗
Figure 4
Figure 4. Figure 4: Visualizing the correlation between reconstruction error and geometric ambigu￾ity manifold. Structural artifacts mainly occur at depth boundaries, and our extracted 2D geometric ambiguity prior aligns closely with these high-error regions, effectively guiding updates to geometrically ambiguous areas. extracted 2D prior. To address the localization challenge, we explicitly detect these visibility transition… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison between Omni-Scene [49] and FocusGS. Six rendered surround views cover the full 360^\circ panorama with about 15\% adjacent-view overlap. Red boxes mark overlap regions, while white dashed regions indicate cases where Fo￾cusGS produces sharper and more complete reconstructions [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Gaussian number of different methods [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Failure cases of our FocusGS [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 11 linked inside Pith

  1. [1]

    In: CVPR

    Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR. pp. 11621–11631 (2020)

    2020

  2. [2]

    In: CVPR

    Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: CVPR. pp. 130–141 (2023)

    2023

  3. [3]

    In: CVPR

    Cao, A.Q., De Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: CVPR. pp. 3991–4001 (2022)

    2022

  4. [4]

    In: ICCV

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650– 9660 (2021)

    2021

  5. [5]

    Charatan,D.,Li,S.L.,Tagliasacchi,A.,Sitzmann,V.:pixelsplat:3dgaussiansplats fromimagepairsforscalablegeneralizable3dreconstruction.In:CVPR.pp.19457– 19467 (2024)

    2024

  6. [6]

    In: ICCV (2021)

    Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV (2021)

    2021

  7. [7]

    IEEE TVCG (2024)

    Chen, D., Li, H., Ye, W., Wang, Y., Xie, W., Zhai, S., Wang, N., Liu, H., Bao, H., Zhang, G.: Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. IEEE TVCG (2024)

    2024

  8. [8]

    In: ECCV

    Chen, Y., Xu, H., Zheng, C., Zhuang, B., Pollefeys, M., Geiger, A., Cham, T.J., Cai, J.: Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In: ECCV. pp. 370–386 (2024)

    2024

  9. [9]

    In: ECCV

    Chen, Z., Ye, M., Xu, S., Cao, T., Chen, Q.: Ppad: Iterative interactions of pre- diction and planning for end-to-end autonomous driving. In: ECCV. pp. 239–256. Springer (2024)

    2024

  10. [10]

    com/traveller59/spconv(2022)

    Contributors, S.: Spconv: Spatially sparse convolution library.https://github. com/traveller59/spconv(2022)

    2022

  11. [11]

    In: CVPR

    Du, Y., Smith, C., Tewari, A., Sitzmann, V.: Learning to render novel views from wide-baseline stereo pairs. In: CVPR. pp. 4970–4980 (2023)

    2023

  12. [12]

    In: ICCV

    Gan, W., Liu, F., Xu, H., Mo, N., Yokoya, N.: Gaussianocc: Fully self-supervised and efficient 3d occupancy estimation with gaussian splatting. In: ICCV. pp. 28980–28990 (2025)

    2025

  13. [13]

    In: CVPR

    Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: CVPR. pp. 17853– 17862 (2023)

    2023

  14. [14]

    arXiv:2112.11790 (2021)

    Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multi- camera 3d object detection in bird-eye-view. arXiv:2112.11790 (2021)

    Pith/arXiv arXiv 2021

  15. [15]

    In: ECCV

    Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In: ECCV. pp. 376–

  16. [16]

    arXiv preprint arXiv:2311.13549 (2023)

    Jia, F., Mao, W., Liu, Y., Zhao, Y., Wen, Y., Zhang, C., Zhang, X., Wang, T.: Adriver-i: A general world model for autonomous driving. arXiv preprint arXiv:2311.13549 (2023)

    Pith/arXiv arXiv 2023

  17. [17]

    In: ICCV

    Jiang, B., Chen, S., Xu, Q., Liao, B., Chen, J., Zhou, H., Zhang, Q., Liu, W., Huang,C.,Wang,X.:Vad:Vectorizedscenerepresentationforefficientautonomous driving. In: ICCV. pp. 8340–8350 (2023)

    2023

  18. [18]

    In: CVPR

    Jiang,H.,Liu,L.,Cheng,T.,Wang,X.,Lin,T.,Su,Z.,Liu,W.,Wang,X.:Gausstr: Foundation model-aligned gaussian transformer for self-supervised 3d spatial un- derstanding. In: CVPR. pp. 11960–11970 (2025) FocusGS 17

    2025

  19. [19]

    ACM TOG44(6), 1–16 (2025)

    Jiang, L., Mao, Y., Xu, L., Lu, T., Ren, K., Jin, Y., Xu, X., Yu, M., Pang, J., Zhao, F., et al.: Anysplat: Feed-forward 3d gaussian splatting from unconstrained views. ACM TOG44(6), 1–16 (2025)

    2025

  20. [20]

    In: CVPR

    Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geom- etry priors. In: CVPR. pp. 18365–18375 (2022)

    2022

  21. [21]

    ACM TOG42(4), 139–1 (2023)

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG42(4), 139–1 (2023)

    2023

  22. [22]

    arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  23. [23]

    AAAI37(2), 1477– 1485 (2023)

    Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. AAAI37(2), 1477– 1485 (2023)

    2023

  24. [24]

    In: ECCV

    Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotem- poral transformers. In: ECCV. pp. 1–18 (2022)

    2022

  25. [25]

    In: CVPR

    Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR. pp. 7824–7833 (2022)

    2022

  26. [26]

    In: ICRA

    Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfu- sion: Multi-task multi-sensor fusion with unified bird’s-eye view representation. In: ICRA. pp. 2774–2781 (2023)

    2023

  27. [27]

    arXiv preprint arXiv:2412.09043 (2024)

    Lu, H., Xu, T., Zheng, W., Zhang, Y., Zhan, W., Du, D., Tomizuka, M., Keutzer, K., Chen, Y.: Drivingrecon: Large 4d gaussian reconstruction model for au- tonomous driving. arXiv preprint arXiv:2412.09043 (2024)

    Pith/arXiv arXiv 2024

  28. [28]

    In: ECCV

    Lu, J., Huang, Z., Yang, Z., Zhang, J., Zhang, L.: Wovogen: World volume-aware diffusion for controllable multi-camera driving scene generation. In: ECCV. pp. 329–345 (2024)

    2024

  29. [29]

    ACM TOG38(4), 1–14 (2019)

    Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescrip- tive sampling guidelines. ACM TOG38(4), 1–14 (2019)

    2019

  30. [30]

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.:Nerf:Representingscenesasneuralradiancefieldsforviewsynthesis.In:ECCV. pp. 405–421 (2020)

    2020

  31. [31]

    In: CVPR

    Min, C., Zhao, D., Xiao, L., Zhao, J., Xu, X., Zhu, Z., Jin, L., Li, J., Guo, Y., Xing, J., et al.: Driveworld: 4d pre-trained scene understanding via world models for autonomous driving. In: CVPR. pp. 15522–15533 (2024)

    2024

  32. [32]

    NeurIPS37, 56828–56858 (2024)

    Ren, J., Xie, C., Mirzaei, A., Kreis, K., Liu, Z., Torralba, A., Fidler, S., Kim, S.W., Ling, H., et al.: L4gm: Large 4d gaussian reconstruction model. NeurIPS37, 56828–56858 (2024)

    2024

  33. [33]

    arXiv preprint arXiv:2403.17898 (2024)

    Ren, K., Jiang, L., Lu, T., Yu, M., Xu, L., Ni, Z., Dai, B.: Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898 (2024)

    Pith/arXiv arXiv 2024

  34. [34]

    NeurIPS 37, 97670–97698 (2024)

    Ren, X., Lu, Y., Liang, H., Wu, Z., Ling, H., Chen, M., Fidler, S., Williams, F., Huang, J.: Scube: Instant large-scale scene reconstruction using voxsplats. NeurIPS 37, 97670–97698 (2024)

    2024

  35. [35]

    arXiv preprint arXiv:2511.04595 (2025)

    Shi, C., Shi, S., Lyu, X., Liu, C., Sheng, K., Zhang, B., Jiang, L.: Unisplat: Unified spatio-temporal fusion via 3d latent scaffolds for dynamic driving scene reconstruc- tion. arXiv preprint arXiv:2511.04595 (2025)

    arXiv 2025

  36. [36]

    NeurIPS 34, 19313–19325 (2021) 18 G

    Sitzmann, V., Rezchikov, S., Freeman, B., Tenenbaum, J., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. NeurIPS 34, 19313–19325 (2021) 18 G. Wang et al

    2021

  37. [37]

    arXiv preprint arXiv:2504.00437 (2025)

    Song, Q., Li, C., Lin, H., Peng, S., Huang, R.: Adgaussian: Generalizable gaus- sian splatting for autonomous driving with multi-modal inputs. arXiv preprint arXiv:2504.00437 (2025)

    arXiv 2025

  38. [38]

    In: CVPR

    Suhail, M., Esteves, C., Sigal, L., Makadia, A.: Light field neural rendering. In: CVPR. pp. 8269–8279 (2022)

    2022

  39. [39]

    In: CVPR

    Tang, P., Wang, Z., Wang, G., Zheng, J., Ren, X., Feng, B., Ma, C.: Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy pre- diction. In: CVPR. pp. 15035–15044 (2024)

    2024

  40. [40]

    AAAI 39(7), 7374–7382 (2025)

    Tian, Q., Tan, X., Xie, Y., Ma, L.: Drivingforward: Feed-forward 3d gaussian splatting for driving scene reconstruction from flexible surround-view input. AAAI 39(7), 7374–7382 (2025)

    2025

  41. [41]

    In: NeurIPS

    Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. In: NeurIPS. vol. 36 (2024)

    2024

  42. [42]

    In: ECCV (2024)

    Wang, G., Wang, Z., Tang, P., Zheng, J., Ren, X., Feng, B., Ma, C.: Occgen: Gen- erative multi-modal 3d occupancy prediction for autonomous driving. In: ECCV (2024)

    2024

  43. [43]

    In: CVPR

    Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin- Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: CVPR. pp. 4690–4699 (2021)

    2021

  44. [44]

    arXiv preprint arXiv:2509.19297 (2025)

    Wang, W., Chen, Y., Zhang, Z., Liu, H., Wang, H., Feng, Z., Qin, W., Zhu, Z., Chen, D.Y., Zhuang, B.: Volsplat: Rethinking feed-forward 3d gaussian splatting with voxel-aligned prediction. arXiv preprint arXiv:2509.19297 (2025)

    Pith/arXiv arXiv 2025

  45. [45]

    In: ECCV

    Wang, X., Zhu, Z., Huang, G., Chen, X., Zhu, J., Lu, J.: Drivedreamer: Towards real-world-drive world models for autonomous driving. In: ECCV. pp. 55–72 (2024)

    2024

  46. [46]

    arXiv:2401.09985 (2024)

    Wang, X., Zhu, Z., Huang, G., Wang, B., Chen, X., Lu, J.: Worlddreamer: To- wards general world models for video generation via predicting masked tokens. arXiv:2401.09985 (2024)

    Pith/arXiv arXiv 2024

  47. [47]

    In: ICCV

    Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: ICCV. pp. 17850–17859 (2023)

    2023

  48. [48]

    IEEE TIP13(4), 600–612 (2004)

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP13(4), 600–612 (2004)

    2004

  49. [49]

    In: CVPR

    Wei, D., Li, Z., Liu, P.: Omni-scene: Omni-gaussian representation for ego-centric sparse-view scene reconstruction. In: CVPR. pp. 22317–22327 (2025)

    2025

  50. [50]

    arXiv:2409.03272 (2024)

    Wei, J., Yuan, S., Li, P., Hu, Q., Gan, Z., Ding, W.: Occllama: An occupancy- language-action generative world model for autonomous driving. arXiv:2409.03272 (2024)

    Pith/arXiv arXiv 2024

  51. [51]

    In: ICCV

    Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi- camera 3d occupancy prediction for autonomous driving. In: ICCV. pp. 21729– 21740 (2023)

    2023

  52. [52]

    Computational Visual Media10(4), 613–642 (2024)

    Wu, T., Yuan, Y.J., Zhang, L.X., Yang, J., Cao, Y.P., Yan, L.Q., Gao, L.: Recent advances in 3d gaussian splatting. Computational Visual Media10(4), 613–642 (2024)

    2024

  53. [53]

    In: CVPR

    Xu, H., Chen, A., Chen, Y., Sakaridis, C., Zhang, Y., Pollefeys, M., Geiger, A., Yu, F.: Murf: multi-baseline radiance fields. In: CVPR. pp. 20041–20050 (2024)

    2024

  54. [54]

    In: CVPR

    Xu, H., Peng, S., Wang, F., Blum, H., Barath, D., Geiger, A., Pollefeys, M.: Depth- splat: Connecting gaussian splatting and depth. In: CVPR. pp. 16453–16463 (2025)

    2025

  55. [55]

    In: ECCV

    Yan, Y., Lin, H., Zhou, C., Wang, W., Sun, H., Zhan, K., Lang, X., Zhou, X., Peng, S.: Street gaussians: Modeling dynamic urban scenes with gaussian splatting. In: ECCV. pp. 156–173 (2024) FocusGS 19

    2024

  56. [56]

    In: ICLR

    Yang, J., Huang, J., Ivanovic, B., Chen, Y., Wang, Y., Li, B., You, Y., Sharma, A., Igl, M., Karkus, P., et al.: Storm: Spatio-temporal reconstruction model for large-scale outdoor scenes. In: ICLR. vol. 2025, pp. 50446–50465 (2025)

    2025

  57. [57]

    In: CVPR

    Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR. pp. 4578–4587 (2021)

    2021

  58. [58]

    In: CVPR

    Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. In: CVPR. pp. 19447–19456 (2024)

    2024

  59. [59]

    In: CVPR

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586–595 (2018)

    2018

  60. [60]

    In: ICCV

    Zhang, Y., Zhu, Z., Du, D.: Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In: ICCV. pp. 9433–9443 (2023)

    2023

  61. [61]

    Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learn- ingviewsynthesisusingmultiplaneimages.arXivpreprintarXiv:1805.09817(2018)

    Pith/arXiv arXiv 2018

  62. [62]

    In: CVPR

    Zhou, X., Lin, Z., Shan, X., Wang, Y., Sun, D., Yang, M.H.: Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In: CVPR. pp. 21634–21643 (2024)

    2024

  63. [63]

    arXiv preprint arXiv:2503.23463 (2025)

    Zhou, X., Han, X., Yang, F., Ma, Y., Knoll, A.C.: Opendrivevla: Towards end-to- end autonomous driving with large vision language action model. arXiv preprint arXiv:2503.23463 (2025)

    arXiv 2025

  64. [64]

    arXiv preprint arXiv:2506.13757 (2025)

    Zhou, Z., Cai, T., Zhao, S.Z., Zhang, Y., Huang, Z., Zhou, B., Ma, J.: Autovla: A vision-language-action model for end-to-end autonomous driving with adaptive reasoning and reinforcement fine-tuning. arXiv preprint arXiv:2506.13757 (2025)

    Pith/arXiv arXiv 2025

  65. [65]

    arXiv preprint arXiv:2010.04159 (2020)

    Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)

    Pith/arXiv arXiv 2010

  66. [66]

    In: CVPR

    Zhu, Z., Wang, S., Xie, J., Liu, J.j., Wang, J., Yang, J.: Voxelsplat: Dynamic gaussian splatting as an effective loss for occupancy and flow prediction. In: CVPR. pp. 6761–6771 (2025)

    2025

  67. [67]

    In: CVPR

    Zuo, S., Zheng, W., Huang, Y., Zhou, J., Lu, J.: Gaussianworld: Gaussian world model for streaming 3d occupancy prediction. In: CVPR. pp. 6772–6781 (2025)

    2025