Pith. sign in

REVIEW 3 major objections 5 minor 103 references

Decoupling geometry from semantics lets sparse 3D occupancy run several times faster while improving completion accuracy.

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 14:24 UTC pith:TRUJG7Z5

load-bearing objection Clean systems paper that actually moves the sparse-occupancy needle: SCF regression on anchors + geometry-guided fill gives real IoU and 4–6× speed gains over their own SparseOcc and dense baselines. the 3 major comments →

arxiv 2607.04732 v1 pith:TRUJG7Z5 submitted 2026-07-06 cs.CV

SparseOcc++: Geometry-Aware Sparse Latent Representation for Semantic Occupancy Prediction

Pin Tang , Zhongdao Wang , Guoqing Wang , Xiangxuan Ren , Chao Ma This is my paper
classification cs.CV
keywords 3D semantic occupancysparse representationscene completion fieldsigned distanceautonomous drivinggeometry-aware perceptionanchor voxels
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.

Vision-based 3D semantic occupancy for driving must fill in occluded space and label every voxel, yet dense grids waste work on empty air and flattened projections lose fine structure. Earlier fully sparse methods still mix the two jobs by flooding empty voxels with high-dimensional features and then classifying them, which produces waste and geometric noise. SparseOcc++ separates the jobs: it first learns a scene completion field on a small set of sparse anchors that predicts signed distances to object boundaries, then expands only those verified regions before any semantic labeling. Orthogonal planar-and-vertical distances plus discretized bins make the geometry robust to outdoor shapes; a single max-pool propagation turns the field into a compact volume. The result is higher IoU and several-times-faster inference than both its predecessor and dense baselines, showing that geometry-first sparse reasoning can keep structural fidelity without dense cost.

Core claim

Scene completion for occupancy prediction is better cast as regression of a vector-valued scene completion field on sparse anchor voxels than as entangled high-dimensional feature propagation plus voxel-wise classification; once the field is converted by geometry-guided propagation into a compact occupied volume, semantic segmentation can be confined to those verified voxels, simultaneously cutting redundancy and reducing geometric ambiguity.

What carries the argument

The scene completion field (SCF): a conditional, anisotropic signed-distance vector defined on coarse-to-fine sparse anchors, decomposed into planar and vertical components, learned by discretized classification, and expanded by radius-guided max-pooling to produce the occupied set for subsequent sparse semantic prediction.

Load-bearing premise

That a single orthogonal planar-plus-vertical distance predicted from sparse anchors, expanded once by max-pooling of that radius, is enough to recover the true occupied volume of complex outdoor objects including non-convex vegetation and distant thin structures.

What would settle it

Measure geometric IoU and per-class completion accuracy on non-convex vegetation, poles, and far-range thin objects after disabling or altering the orthogonal decomposition and single-iteration max-pool propagation; a clear drop relative to dense or multi-iteration baselines would falsify sufficiency of the SCF representation.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Occupancy pipelines can drop dense 3D operators and still recover complete geometry by first regressing sparse distance fields.
  • Semantic heads need only run on geometrically verified voxels, yielding real-time FPS on multi-camera driving scenes.
  • Anchor-based geometric regression becomes a reusable primitive for other sparse 3D perception tasks that currently mix geometry and semantics.
  • Discretized orthogonal distances offer a practical way to handle outdoor anisotropy without continuous SDF solvers.

Where Pith is reading between the lines

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

  • The same SCF-plus-propagation pattern could be applied to LiDAR or multi-modal fusion by initializing anchors from range data rather than image features alone.
  • Temporal consistency of the SCF across frames may resolve the long-range depth ambiguity the paper notes as a remaining limit.
  • Adaptive rather than fixed orthogonal axes could better capture non-convex shapes the current model oversimplifies.

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

3 major / 5 minor

Summary. The paper proposes SparseOcc++, a fully sparse vision-based 3D semantic occupancy framework that explicitly disentangles scene completion from semantic classification. It introduces sparse anchor voxels generated coarse-to-fine via deformable cross-attention, learns a vector-valued Scene Completion Field (SCF) on them via orthogonal (planar X-Y + vertical Z) decomposition and discretized classification, then applies geometry-guided max-pool propagation (Algorithm 1) to densify only verified occupied regions before a lightweight sparse U-Net performs semantics. Building on the authors’ prior SparseOcc, this yields new SOTA results: +2.3 IoU and 3.9× FPS versus SparseOcc on nuScenes-Occupancy, and 5.9× speedup versus OccFormer on SemanticKITTI, with supporting ablations (Tables III–XI) and qualitative gains on small/dynamic objects.

Significance. If the empirical gains hold, the work is significant for real-time autonomous-driving perception: it demonstrates that a geometry-first sparse latent representation can simultaneously improve structural fidelity and reduce computation relative to both dense voxel methods and prior entangled sparse methods. Strengths include extensive component ablations on public benchmarks, explicit disclosure of the non-convex limitation (Sec. VI), and reproducible algorithmic details (Algorithms 1–2, free parameters listed). The orthogonal-SCF + guided-propagation design is a clean, falsifiable contribution that others can build upon or refute.

major comments (3)
  1. Sec. IV-C / Algorithm 2: SCF ground-truth is generated by iterative average-pooling rather than a true Euclidean Distance Transform. While efficient, the homogeneous-window test only guarantees a lower bound on distance and can under-estimate for non-axis-aligned or non-convex surfaces; the paper should quantify the approximation error (e.g., mean absolute deviation versus EDT on a subset of SemanticKITTI) or show that the discretized bins remain robust to it, because the entire regression target rests on this supervision.
  2. Tables I–II and Sec. V-D: The SOTA IoU/FPS claims are reported from single training runs without error bars or multi-seed statistics. Given that free parameters (s_X-Y_max, s_Z_max, top-k ratios) and random anchor initialization exist, a modest multi-run evaluation (or at least variance on the key SparseOcc++ vs SparseOcc delta) is needed to confirm that the 2.3-point IoU and 3.9× speed-up are stable rather than seed-dependent.
  3. Sec. IV-C and Algorithm 1: Geometry-guided propagation expands each anchor by a single isotropic max-pool of radius |ŝ|. For thin or elongated structures (poles, distant vegetation) whose true extent is anisotropic beyond the orthogonal decomposition, this can still produce over- or under-completion; Table X shows sensitivity to (s_max, iterations), yet the paper does not measure the fraction of false-positive voxels introduced by the max-pool step itself. A short diagnostic (precision of the completed mask before the U-Net) would strengthen the central efficiency claim.
minor comments (5)
  1. Table II title contains the typo “occpancy”; correct to “occupancy”.
  2. Fig. 5 caption and surrounding text: “directional distance” is used interchangeably with SCF; a single consistent term would improve readability.
  3. Eq. (9) and the subsequent sign convention: positive = empty, negative = occupied is opposite to the classic SDF convention; a brief remark would avoid confusion for graphics readers.
  4. Sec. V-C: the exact top-k retention ratios {1/5, 1/10} and the number of DCA sampling points are stated only in prose; listing them in a hyper-parameter table would aid reproducibility.
  5. Fig. 8 y-axes are unlabeled in the text description; ensure the published figure has clear “Inference Memory (MB)” and “#Voxels” labels.

Circularity Check

0 steps flagged

No significant circularity; empirical SOTA claims rest on external benchmarks and ablations of new modules, with only routine self-citation of the preliminary SparseOcc baseline.

full rationale

The paper's central claims are empirical (IoU +2.3 and 3.9x FPS vs SparseOcc on nuScenes; 5.9x vs OccFormer on SemanticKITTI) measured against public datasets and independently published dense/sparse baselines (Tables I-II). SCF is defined as a vector-valued signed-distance field on sparse anchors (Eqs. 7-9), with GT generated by a deterministic iterative avg-pool transform of binary occupancy labels (Algorithm 2); this does not embed the final IoU/mIoU metric into the training target. Geometry-guided propagation (Algorithm 1) and subsequent Sparse U-Net semantics are supervised by standard occupancy losses (Eqs. 13-14). Orthogonal decomposition and discretized learning are design choices ablated in Tables VII-X, not forced by self-citation of a uniqueness result. Self-citation of SparseOcc [27] is present (as expected for an extension) but non-load-bearing: the new geometry-aware modules are explicitly ablated against the entangled baseline, and gains hold vs external methods. No fitted parameter is renamed as a prediction, no ansatz is smuggled via prior self-work, and no derivation reduces by construction to its inputs. The acknowledged limitation on non-convex geometry (Sec. VI) is orthogonal to circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 2 invented entities

The central empirical claim rests on a small set of domain assumptions about outdoor sparsity and geometric anisotropy, a handful of hand-chosen free parameters that control anchor density and propagation radius, and two invented representational entities (SCF and sparse anchor voxels) that have no independent physical existence outside the method.

free parameters (4)
  • s_X-Y_max (max planar SCF distance) = 10
    Hard-coded to 10 voxels; controls the largest horizontal propagation radius and the number of classification bins. Chosen by ablation (Table X) rather than derived.
  • s_Z_max (max vertical SCF distance) = 5
    Hard-coded to 5 voxels; same role for the height axis. Chosen by ablation.
  • top-k retention ratios for pyramidal anchor generation = 1/5, 1/10
    Keep top 1/5 then 1/10 of candidates at successive resolution levels. Directly determines how many anchors enter SCF learning; selected for efficiency–coverage trade-off.
  • number of SCF completion iterations = 1
    Set to 1 after ablation showed that two iterations can degrade mIoU via over-propagation.
axioms (3)
  • domain assumption Outdoor driving scenes are sufficiently sparse that a small set of dynamically selected anchors plus local distance-based filling recovers the true occupied volume.
    Stated in Sec. I and IV-B; underpins the entire sparse design. Supported by the 67–80 % empty-voxel statistics but not proved for all geometries.
  • ad hoc to paper Geometric anisotropy of outdoor objects is adequately captured by independent planar (X-Y) and vertical (Z) signed distances.
    Introduced in Sec. IV-C and Fig. 5; the orthogonal decomposition is a modeling choice, not a theorem.
  • domain assumption Discretizing continuous signed distances into classification bins stabilizes training and is preferable to direct regression.
    Borrowed from depth-estimation practice (cited) and validated by ablation (Table IX).
invented entities (2)
  • Scene Completion Field (SCF) no independent evidence
    purpose: Vector-valued signed-distance function defined on sparse anchors that encodes anisotropic geometric extent for subsequent propagation.
    Defined in Eq. (8)–(9); differs from classical SDF by being conditional on image features, anisotropic, and used only for completion rather than rendering. No independent physical measurement exists outside the method.
  • Sparse anchor voxels with geometric attribute s_i no independent evidence
    purpose: Dynamically generated proposals that carry both feature embeddings and SCF distances, analogous to anchor boxes in detection.
    Introduced in Eq. (7) and Sec. IV-B; the geometric attribute s_i is new relative to prior sparse occupancy representations.

pith-pipeline@v1.1.0-grok45 · 32219 in / 2915 out tokens · 28340 ms · 2026-07-11T14:24:22.098860+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of SparseOcc++: Geometry-Aware Sparse Latent Representation for Semantic Occupancy Prediction." pith.science (2026). https://pith.science/paper/TRUJG7Z5

@misc{pith2026260704732,
  author       = {Pith},
  title        = {Pith review of: SparseOcc++: Geometry-Aware Sparse Latent Representation for Semantic Occupancy Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TRUJG7Z5}},
  note         = {Machine review of arXiv:2607.04732}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Vision-based 3D semantic occupancy prediction is essential for autonomous driving, yet dense voxel representations waste computation on largely empty space, while BEV and TPV projections compromise fine-grained 3D structure. Fully sparse representations offer an attractive alternative, but existing methods, including SparseOcc, entangle scene completion with semantic prediction by indiscriminately propagating high-dimensional features into empty regions and applying voxel-wise classification. This creates excessive activations, computational overhead, and geometric ambiguity. We present SparseOcc++, a geometry-aware sparse framework that explicitly decouples scene completion from semantic segmentation. SparseOcc++ reformulates completion as signed-distance regression on sparse anchor voxels through a scene completion field (SCF). To model complex outdoor geometry robustly, it combines orthogonal decomposition with discretized distance learning. A geometry-guided propagation module then converts the SCF into a complete volumetric scene and restricts semantic segmentation to geometrically verified regions. Experiments establish new state of the art: SparseOcc++ improves IoU by 2.3 points and is 3.9x faster than SparseOcc on nuScenes, while achieving a 5.9x speedup over OccFormer on SemanticKITTI.

Figures

Figures reproduced from arXiv: 2607.04732 by Chao Ma, Guoqing Wang, Pin Tang, Xiangxuan Ren, Zhongdao Wang.

Figure 1
Figure 1. Figure 1: (a) Prior approaches primarily rely on dense [1], [2], BEV [3], [4], or TPV [5] representations. In contrast, we explore [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed SparseOcc. (a) Images captured by monocular or surrounding cameras are first passed to a 2D encoder, yielding 2D latent features. Then the latent features are mapped to 3D using the predicted depth map following the LSS [9]. (b) SparseOcc adopts a sparse representation for the latent space. Upon this representation, we introduce three key building blocks: a sparse latent propagator… view at source ↗
Figure 3
Figure 3. Figure 3: C. Sparse Feature Pyramid A straightforward approach to completing the scene is to stack the proposed sparse propagator multiple times. However, this necessitates a substantial number of sparse propagators to ensure an adequately large receptive field, which is particularly important for recognizing large objects like ‘truck’ or static elements such as ‘road’. The computational cost is obviously high. To a… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the proposed SparseOcc++. It first generates a set of sparse anchor voxels, which are employed as queries to gather features from images via deformable cross attention. To achieve compact scene completion, a scene completion field (SCF) is learned on these anchor voxels for efficient geometry-guided propagation. Finally, it employs a lightweight sparse U-Net to predict the final fine-grained se… view at source ↗
Figure 5
Figure 5. Figure 5: Visualizations on directional distance. (a) The input image. (b) A single scalar directional distance is insufficient to model the scene due to geometric anisotropy. (c)&(d) The proposed orthogonal decomposition strategy decouples the directional distance into an X–Y planar component and a vertical Z-axis component, providing a closer match to the geometric distribution. s X-Y measures the distance to the … view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results on the SemanticKITTI validation set. The input monocular image is shown on the leftmost side, and the 3D occupancy predictions of the dense baseline OccFormer [40], SparseOcc, SparseOcc++, and the ground truth are then visualized sequentially. The regions highlighted by rectangles show areas with noticeable differences. Best viewed in color and zoomed in. Input C-CONet SparseOcc SparseO… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results of 3D semantic occupancy on the nuScenes-Occupancy validation set. The input multi-view images are shown on the leftmost side, and the occupancy predictions of the dense baseline C-CONet [2], SparseOcc, SparseOcc++, and the ground-truth are then visualized sequentially. Compared to the 3D dense representation based C￾CONet [20], our sparse representation achieves better completion and s… view at source ↗
Figure 8
Figure 8. Figure 8: Efficiency analysis when scaling up the 3D repre￾sentation on SemanticKITTI [41] validation set. The left axis represents the inference GPU memory. The right axis denotes the number of voxels. The 3D downsampling ratio is considered between the ground-truth and the LSS output. The number of non-empty voxels is measured using max￾donwsampled ground-truth. the explicit volumetric structure. Extensive experim… 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

103 extracted references · 15 linked inside Pith

  1. [1]

    V oxnet: A 3d convolutional neural network for real-time object recognition,

    D. Maturana and S. Scherer, “V oxnet: A 3d convolutional neural network for real-time object recognition,” inIROS, 2015

    2015

  2. [2]

    Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception,

    X. Wang, Z. Zhu, W. Xu, Y . Zhang, Y . Wei, X. Chi, Y . Ye, D. Du, J. Lu, and X. Wang, “Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception,” inICCV, 2023

    2023

  3. [3]

    Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers,

    Z. Li, W. Wang, H. Li, E. Xie, C. Sima, T. Lu, Y . Qiao, and J. Dai, “Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers,” inECCV, 2022

    2022

  4. [4]

    Bevdet: High-performance multi-camera 3d object detection in bird-eye-view,

    J. Huang, G. Huang, Z. Zhu, and D. Du, “Bevdet: High-performance multi-camera 3d object detection in bird-eye-view,”arXiv:2112.11790, 2021

    Pith/arXiv arXiv 2021

  5. [5]

    Tri-perspective view for vision-based 3d semantic occupancy prediction,

    Y . Huang, W. Zheng, Y . Zhang, J. Zhou, and J. Lu, “Tri-perspective view for vision-based 3d semantic occupancy prediction,” inCVPR, 2023. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 16

    2023

  6. [6]

    Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving,

    X. Jia, Y . Gao, L. Chen, J. Yan, P. L. Liu, and H. Li, “Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving,” inCVPR, 2023

    2023

  7. [7]

    Planning-oriented autonomous driving,

    Y . Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wanget al., “Planning-oriented autonomous driving,” in CVPR, 2023

    2023

  8. [8]

    Genad: Generative end-to-end autonomous driving,

    W. Zheng, R. Song, X. Guo, C. Zhang, and L. Chen, “Genad: Generative end-to-end autonomous driving,” inECCV, 2024

    2024

  9. [9]

    Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,

    J. Philion and S. Fidler, “Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,” inECCV, 2020

    2020

  10. [10]

    Bevdepth: Acquisition of reliable depth for multi-view 3d object detection,

    Y . Li, Z. Ge, G. Yu, J. Yang, Z. Wang, Y . Shi, J. Sun, and Z. Li, “Bevdepth: Acquisition of reliable depth for multi-view 3d object detection,”arXiv:2206.10092, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” in CVPR, 2019

    2019

  12. [12]

    Alleviating foreground sparsity for semi-supervised monocular 3d object detection,

    W. Zhang, D. Liu, C. Ma, and W. Cai, “Alleviating foreground sparsity for semi-supervised monocular 3d object detection,” inWACV, 2024

    2024

  13. [13]

    Unidistill: A universal cross-modality knowledge distillation framework for 3d object detec- tion in bird’s-eye view,

    S. Zhou, W. Liu, C. Hu, S. Zhou, and C. Ma, “Unidistill: A universal cross-modality knowledge distillation framework for 3d object detec- tion in bird’s-eye view,” inCVPR, 2023

    2023

  14. [14]

    Bi-stream knowledge transfer for semi-supervised 3d point cloud object detection,

    J. Zheng, P. Tang, X. Ren, Z. Wang, and C. Ma, “Bi-stream knowledge transfer for semi-supervised 3d point cloud object detection,” inICRA, 2025

    2025

  15. [15]

    Litefusion: Taming 3d object detectors from vision-based to multi-modal with minimal adaptation,

    X. Ren, Z. Wang, P. Tang, G. Wang, J. Zheng, and C. Ma, “Litefusion: Taming 3d object detectors from vision-based to multi-modal with minimal adaptation,”arXiv:2512.20217, 2025

    arXiv 2025

  16. [16]

    Transfusion: Robust lidar-camera fusion for 3d object detection with transformers,

    X. Bai, Z. Hu, X. Zhu, Q. Huang, Y . Chen, H. Fu, and C.-L. Tai, “Transfusion: Robust lidar-camera fusion for 3d object detection with transformers,” inCVPR, 2022

    2022

  17. [17]

    Second: Sparsely embedded convolutional detection,

    Y . Yan, Y . Mao, and B. Li, “Second: Sparsely embedded convolutional detection,”Sensors, 2018

    2018

  18. [18]

    Rclane: Relay chain prediction for lane detection,

    S. Xu, X. Cai, B. Zhao, L. Zhang, H. Xu, Y . Fu, and X. Xue, “Rclane: Relay chain prediction for lane detection,” inECCV, 2022

    2022

  19. [19]

    Prototransfer: Cross-modal prototype transfer for point cloud segmentation,

    P. Tang, H.-M. Xu, and C. Ma, “Prototransfer: Cross-modal prototype transfer for point cloud segmentation,” inICCV, 2023

    2023

  20. [20]

    2dpass: 2d priors assisted semantic segmentation on lidar point clouds,

    X. Yan, J. Gao, C. Zheng, C. Zheng, R. Zhang, S. Cui, and Z. Li, “2dpass: 2d priors assisted semantic segmentation on lidar point clouds,” inECCV, 2022

    2022

  21. [22]

    Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase,

    Y . Liu, R. Chen, X. Li, L. Kong, Y . Yang, Z. Xia, Y . Bai, X. Zhu, Y . Ma, Y . Liet al., “Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase,” inICCV, 2023

    2023

  22. [23]

    Mseg3d: Multi-modal 3d semantic segmentation for autonomous driving,

    J. Li, H. Dai, H. Han, and Y . Ding, “Mseg3d: Multi-modal 3d semantic segmentation for autonomous driving,” inCVPR, 2023

    2023

  23. [24]

    Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving,

    Y . Wei, L. Zhao, W. Zheng, Z. Zhu, J. Zhou, and J. Lu, “Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving,” in CVPR, 2023

    2023

  24. [25]

    Octreeocc: Efficient and multi-granularity occupancy prediction using octree queries,

    Y . Lu, X. Zhu, T. Wang, and Y . Ma, “Octreeocc: Efficient and multi-granularity occupancy prediction using octree queries,” arXiv:2312.03774, 2023

    Pith/arXiv arXiv 2023

  25. [26]

    Opus: occupancy prediction using a sparse set,

    J. Wang, Z. Liu, Q. Meng, L. Yan, K. Wang, J. Yang, W. Liu, Q. Hou, and M.-M. Cheng, “Opus: occupancy prediction using a sparse set,” in NeurIPS, 2024

    2024

  26. [27]

    Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction,

    P. Tang, Z. Wang, G. Wang, J. Zheng, X. Ren, B. Feng, and C. Ma, “Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction,” inCVPR, 2024

    2024

  27. [28]

    Fully sparse 3d panoptic occupancy prediction,

    H. Liu, H. Wang, Y . Chen, Z. Yang, J. Zeng, L. Chen, and L. Wang, “Fully sparse 3d panoptic occupancy prediction,” inECCV, 2024

    2024

  28. [29]

    Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin,

    Z. Yu, C. Shu, J. Deng, K. Lu, Z. Liu, J. Yu, D. Yang, H. Li, and Y . Chen, “Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin,”arXiv:2311.12058, 2023

    Pith/arXiv arXiv 2023

  29. [30]

    Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction,

    Y . Huang, W. Zheng, Y . Zhang, J. Zhou, and J. Lu, “Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction,” arXiv:2405.17429, 2024

    Pith/arXiv arXiv 2024

  30. [31]

    Probabilistic gaussian superposition for efficient 3d occupancy prediction,

    Y . Huang, A. Thammatadatrakoon, W. Zheng, Y . Zhang, D. Du, and J. Lu, “Probabilistic gaussian superposition for efficient 3d occupancy prediction,”arXiv:2412.04384, 2024

    Pith/arXiv arXiv 2024

  31. [32]

    Fb-occ: 3d occupancy prediction based on forward-backward view transformation,

    Z. Li, Z. Yu, D. Austin, M. Fang, S. Lan, J. Kautz, and J. M. Alvarez, “Fb-occ: 3d occupancy prediction based on forward-backward view transformation,”arXiv:2307.01492, 2023

    Pith/arXiv arXiv 2023

  32. [33]

    Scene as occupancy,

    W. Tong, C. Sima, T. Wang, L. Chen, S. Wu, H. Deng, Y . Gu, L. Lu, P. Luo, D. Linet al., “Scene as occupancy,” inICCV, 2023

    2023

  33. [34]

    Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving,

    X. Tian, T. Jiang, L. Yun, Y . Wang, Y . Wang, and H. Zhao, “Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving,”arXiv:2304.14365, 2023

    Pith/arXiv arXiv 2023

  34. [35]

    Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,

    Z. Liu, H. Tang, A. Amini, X. Yang, H. Mao, D. Rus, and S. Han, “Bevfusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,”arXiv:2205.13542, 2022

    Pith/arXiv arXiv 2022

  35. [36]

    Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction,

    S. Zuo, W. Zheng, Y . Huang, J. Zhou, and J. Lu, “Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction,” arXiv:2308.16896, 2023

    Pith/arXiv arXiv 2023

  36. [37]

    Spconv: Spatially sparse convolution library,

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

    2022

  37. [38]

    Faster R-CNN: Towards real-time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards real-time object detection with region proposal networks,” inNeurIPS, 2015

    2015

  38. [39]

    Deepsdf: Learning continuous signed distance functions for shape representation,

    J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove, “Deepsdf: Learning continuous signed distance functions for shape representation,” inCVPR, 2019

    2019

  39. [40]

    Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction,

    Y . Zhang, Z. Zhu, and D. Du, “Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction,” inICCV, 2023

    2023

  40. [41]

    Semantickitti: A dataset for semantic scene understanding of lidar sequences,

    J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall, “Semantickitti: A dataset for semantic scene understanding of lidar sequences,” inICCV, 2019

    2019

  41. [42]

    Monoscene: Monocular 3d semantic scene completion,

    A.-Q. Cao and R. de Charette, “Monoscene: Monocular 3d semantic scene completion,” inCVPR, 2022

    2022

  42. [43]

    Nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “Nuscenes: A multimodal dataset for autonomous driving,” inCVPR, 2020

    2020

  43. [44]

    Scalability in perception for autonomous driving: Waymo open dataset,

    P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V . Patnaik, P. Tsui, J. Guo, Y . Zhou, Y . Chai, B. Caine, V . K. Vasudevan, W. Han, J. Ngiam, H. Zhao, A. Timofeev, S. Ettinger, M. Krivokon, A. Gao, A. Joshi, Y . Zhang, J. Shlens, Z. Chen, and D. Anguelov, “Scalability in perception for autonomous driving: Waymo open dataset,” inCVPR, 2020

    2020

  44. [45]

    Fastocc: Accelerating 3d occupancy prediction by fusing the 2d bird’s-eye view and perspective view,

    J. Hou, X. Li, W. Guan, G. Zhang, D. Feng, Y . Du, X. Xue, and J. Pu, “Fastocc: Accelerating 3d occupancy prediction by fusing the 2d bird’s-eye view and perspective view,” inICRA, 2024

    2024

  45. [46]

    Occrwkv: Rethinking efficient 3d semantic occupancy prediction with linear complexity,

    J. Wang, W. Yin, X. Long, X. Zhang, Z. Xing, X. Guo, and Q. Zhang, “Occrwkv: Rethinking efficient 3d semantic occupancy prediction with linear complexity,” inICRA, 2025

    2025

  46. [47]

    Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation,

    Y . Wang, Y . Chen, X. Liao, L. Fan, and Z. Zhang, “Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation,” inCVPR, 2024

    2024

  47. [48]

    V oxelsplat: Dynamic gaussian splatting as an effective loss for occupancy and flow prediction,

    Z. Zhu, S. Wang, J. Xie, J.-j. Liu, J. Wang, and J. Yang, “V oxelsplat: Dynamic gaussian splatting as an effective loss for occupancy and flow prediction,” inCVPR, 2025

    2025

  48. [49]

    Gs-occ3d: Scaling vision-only occupancy reconstruction with gaussian splatting,

    B. Ye, M. Qin, S. Zhang, M. Gong, S. Zhu, H. Zhao, and H. Zhao, “Gs-occ3d: Scaling vision-only occupancy reconstruction with gaussian splatting,” inICCV, 2025

    2025

  49. [50]

    Gauss- render: Learning 3d occupancy with gaussian rendering,

    L. Chambon, E. Zablocki, A. Boulch, M. Chen, and M. Cord, “Gauss- render: Learning 3d occupancy with gaussian rendering,” inICCV, 2025

    2025

  50. [51]

    Gaussianworld: Gaussian world model for streaming 3d occupancy prediction,

    S. Zuo, W. Zheng, Y . Huang, J. Zhou, and J. Lu, “Gaussianworld: Gaussian world model for streaming 3d occupancy prediction,” in CVPR, 2025

    2025

  51. [52]

    Visionpad: A vision-centric pre-training paradigm for autonomous driving,

    H. Zhang, W. Zhou, Y . Zhu, X. Yan, J. Gao, D. Bai, Y . Cai, B. Liu, S. Cui, and Z. Li, “Visionpad: A vision-centric pre-training paradigm for autonomous driving,” inCVPR, 2025

    2025

  52. [53]

    Occmamba: Semantic occupancy prediction with state space models,

    H. Li, Y . Hou, X. Xing, Y . Ma, X. Sun, and Y . Zhang, “Occmamba: Semantic occupancy prediction with state space models,” inCVPR, 2025

    2025

  53. [54]

    Alocc: Adaptive lifting-based 3d semantic occupancy and cost volume-based flow predictions,

    D. Chen, J. Fang, W. Han, X. Cheng, J. Yin, C. Xu, F. S. Khan, and J. Shen, “Alocc: Adaptive lifting-based 3d semantic occupancy and cost volume-based flow predictions,” inICCV, 2025

    2025

  54. [55]

    Dvgt: Driving visual geometry transformer,

    S. Zuo, Z. Xie, W. Zheng, S. Xu, F. Li, S. Jiang, L. Chen, Z.- X. Yang, and J. Lu, “Dvgt: Driving visual geometry transformer,” arXiv:2512.16919, 2025

    Pith/arXiv arXiv 2025

  55. [56]

    Pop-3d: Open-vocabulary 3d occupancy prediction from images,

    A. V obecky, O. Sim ´eoni, D. Hurych, S. Gidaris, A. Bursuc, P. P ´erez, and J. Sivic, “Pop-3d: Open-vocabulary 3d occupancy prediction from images,” inNeurIPS, 2023

    2023

  56. [57]

    Veon: V ocabulary-enhanced occupancy prediction,

    J. Zheng, P. Tang, Z. Wang, G. Wang, X. Ren, B. Feng, and C. Ma, “Veon: V ocabulary-enhanced occupancy prediction,” inECCV, 2024

    2024

  57. [58]

    Language driven occupancy prediction,

    Z. Yu, B. Pang, L. Liu, R. Zhang, Q. Li, S.-Y . Cao, M. Luo, M. Chen, S. Yang, and H.-L. Shen, “Language driven occupancy prediction,” in ICCV, 2025

    2025

  58. [59]

    Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation,

    Y . Wang, Y . Chen, X. Liao, L. Fan, and Z. Zhang, “Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation,” inCVPR, 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 17

    2024

  59. [60]

    Geocc: Geometrically enhanced 3d occupancy network with implicit-explicit depth fusion and contextual self-supervision,

    X. Tan, W. Wu, Z. Zhang, C. Fan, Y . Peng, Z. Zhang, Y . Xie, and L. Ma, “Geocc: Geometrically enhanced 3d occupancy network with implicit-explicit depth fusion and contextual self-supervision,”IEEE TITS, 2025

    2025

  60. [61]

    Selfocc: Self- supervised vision-based 3d occupancy prediction,

    Y . Huang, W. Zheng, B. Zhang, J. Zhou, and J. Lu, “Selfocc: Self- supervised vision-based 3d occupancy prediction,” inCVPR, 2024

    2024

  61. [62]

    Gaussianocc: Fully self-supervised and efficient 3d occupancy estimation with gaussian splatting,

    W. Gan, F. Liu, H. Xu, N. Mo, and N. Yokoya, “Gaussianocc: Fully self-supervised and efficient 3d occupancy estimation with gaussian splatting,” inICCV, 2025

    2025

  62. [63]

    Gausstr: Foundation model-aligned gaussian transformer for self-supervised 3d spatial understanding,

    H. Jiang, L. Liu, T. Cheng, X. Wang, T. Lin, Z. Su, W. Liu, and X. Wang, “Gausstr: Foundation model-aligned gaussian transformer for self-supervised 3d spatial understanding,” inCVPR, 2025

    2025

  63. [64]

    Lidar-camera continuous fusion in voxelized grid for semantic scene completion,

    Z. Lu, B. Cao, and Q. Hu, “Lidar-camera continuous fusion in voxelized grid for semantic scene completion,”IEEE TCSVT, 2024

    2024

  64. [65]

    Self-supervised sparse sensor fusion for long range per- ception,

    E. Palladin, S. Brucker, F. Ghilotti, P. Narayanan, M. Bijelic, and F. Heide, “Self-supervised sparse sensor fusion for long range per- ception,” inICCV, 2025

    2025

  65. [66]

    Riocc: Efficient cross-modal fusion transformer with collaborative feature refinement for 3d semantic occupancy prediction,

    B. Fan, X. Li, Y . Zhou, Y . Jiang, J. Tian, and H. Fan, “Riocc: Efficient cross-modal fusion transformer with collaborative feature refinement for 3d semantic occupancy prediction,” inCVPR, 2025

    2025

  66. [67]

    Sgformer: Satellite- ground fusion for 3d semantic scene completion,

    X. Guo, J. Hu, J. Hu, H. Bao, and G. Zhang, “Sgformer: Satellite- ground fusion for 3d semantic scene completion,” inCVPR, 2025

    2025

  67. [68]

    Omninwm: Omniscient driving navigation world models,

    B. Li, Z. Ma, D. Du, B. Peng, Z. Liang, Z. Liu, C. Ma, Y . Jin, H. Zhao, W. Zenget al., “Omninwm: Omniscient driving navigation world models,”arXiv:2510.18313, 2025

    Pith/arXiv arXiv 2025

  68. [69]

    Oc- cgen: Generative multi-modal 3d occupancy prediction for autonomous driving,

    G. Wang, Z. Wang, P. Tang, J. Zheng, X. Ren, B. Feng, and C. Ma, “Oc- cgen: Generative multi-modal 3d occupancy prediction for autonomous driving,” inECCV, 2024

    2024

  69. [70]

    Occworld: Learning a 3d occupancy world model for autonomous driving,

    W. Zheng, W. Chen, Y . Huang, B. Zhang, Y . Duan, and J. Lu, “Occworld: Learning a 3d occupancy world model for autonomous driving,” inECCV, 2024

    2024

  70. [71]

    Scpnet: Semantic scene completion on point cloud,

    Z. Xia, Y . Liu, X. Li, X. Zhu, Y . Ma, Y . Li, Y . Hou, and Y . Qiao, “Scpnet: Semantic scene completion on point cloud,” inCVPR, 2023

    2023

  71. [72]

    V oxformer: Sparse voxel transformer for camera- based 3d semantic scene completion,

    Y . Li, Z. Yu, C. Choy, C. Xiao, J. M. Alvarez, S. Fidler, C. Feng, and A. Anandkumar, “V oxformer: Sparse voxel transformer for camera- based 3d semantic scene completion,” inCVPR, 2023

    2023

  72. [73]

    Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering,

    M. Pan, L. Liu, J. Liu, P. Huang, L. Wang, S. Zhang, S. Xu, Z. Lai, and K. Yang, “Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering,”arXiv:2306.09117, 2023

    Pith/arXiv arXiv 2023

  73. [74]

    Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision,

    M. Pan, J. Liu, R. Zhang, P. Huang, X. Li, H. Xie, B. Wang, L. Liu, and S. Zhang, “Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision,” inICRA, 2024

    2024

  74. [75]

    Occnerf: Self-supervised multi-camera occupancy prediction with neural radiance fields,

    C. Zhang, J. Yan, Y . Wei, J. Li, L. Liu, Y . Tang, Y . Duan, and J. Lu, “Occnerf: Self-supervised multi-camera occupancy prediction with neural radiance fields,”arXiv:2312.09243, 2023

    Pith/arXiv arXiv 2023

  75. [76]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering.”ACM TOG, vol. 42, no. 4, pp. 139–1, 2023

    2023

  76. [77]

    V oxelnet: End-to-end learning for point cloud based 3d object detection,

    Y . Zhou and O. Tuzel, “V oxelnet: End-to-end learning for point cloud based 3d object detection,” inCVPR, 2018

    2018

  77. [78]

    Semantic scene completion from a single depth image,

    S. Song, F. Yu, A. Zeng, A. X. Chang, M. Savva, and T. Funkhouser, “Semantic scene completion from a single depth image,” inCVPR, 2017

    2017

  78. [79]

    Semantic scene completion using local deep implicit functions on lidar data,

    C. B. Rist, D. Emmerichs, M. Enzweiler, and D. M. Gavrila, “Semantic scene completion using local deep implicit functions on lidar data,” IEEE TPAMI, vol. 44, no. 10, pp. 7205–7218, 2021

    2021

  79. [80]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inCVPR, 2016

    2016

  80. [81]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” inCVPR, 2017

    2017

Showing first 80 references.