Pith. sign in

REVIEW 3 major objections 5 minor 65 references

GPOcc++ claims that surface-only geometry predictions can be converted into dense semantic 3D occupancy by placing sparse Gaussian primitives along camera rays, and backs the claim with top results on indoor and outdoor benchmarks.

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 · deepseek-v4-flash

2026-08-02 05:02 UTC pith:22PBLZ2U

load-bearing objection Solid extension of GPOcc, but the headline indoor gains are confounded by the VGGT backbone and the nuScenes claims overreach. the 3 major comments →

arxiv 2607.13481 v1 pith:22PBLZ2U submitted 2026-07-15 cs.CV

GPOcc++: Unified Sparse Gaussian Occupancy Prediction with Visual Geometry Priors

Changqing Zhou , Yueru Luo , Yulan Guo , Bing Wang , Jie Qin , Changhao Chen This is my paper
classification cs.CV
keywords 3D occupancy predictionsemantic scene completionsparse Gaussian representationvisual geometry priorsray-based volumetric samplingmulti-view fusiontemporal fusionembodied AI
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.

This paper tries to establish that the gap between surface-centric visual geometry priors and volumetric occupancy prediction can be closed by a sparse Gaussian representation. Instead of lifting all image features into a dense 3D grid or scattering random anchors, it extends each predicted surface point along its camera ray to create volumetric anchors, each of which predicts a semantic Gaussian primitive. The Gaussian set is pruned by opacity and probabilistically aggregated into voxel occupancy. The extended system, GPOcc++, adds ray-conditioned fusion across multi-view and temporal observations and per-anchor 3D offset prediction. If the results hold, occupancy prediction can directly piggyback on rapidly improving geometry foundation models while staying compact and efficient enough for embodied agents and autonomous driving.

Core claim

The paper's central discovery is that surface-level predictions from visual geometry priors—depth maps or pointmaps—can be turned into volumetric occupancy by ray-based volumetric sampling. For each pixel, K anchors are placed along the camera ray at offsets scaled by a learned predictor, so the Gaussian primitives concentrate on and inside objects rather than in empty space. Opacity pruning removes low-contribution primitives, and a probabilistic Gaussian-to-occupancy aggregation yields semantic voxel predictions. GPOcc++ extends this with Ray-Conditioned Multi-Image Fusion, which injects Plücker ray coordinates into cross-image attention to relate multi-view and temporal observations, and

What carries the argument

The load-bearing mechanism is ray-based volumetric sampling with sparse semantic Gaussians. A predicted surface point at depth d is extended to K anchors at distances d + delta_k along the camera ray, with the offsets generated by a learned scale predictor; each anchor predicts a Gaussian primitive with center, scale, rotation, opacity, and semantic features. The Gaussians are pruned at opacity threshold tau and mapped to voxel occupancy through probabilistic superposition. Two additions carry GPOcc++: Ray-Conditioned Multi-Image Fusion, where each pixel's ray is encoded as a Plücker coordinate and added to image tokens before global attention across frames/views, and Offset-Guided Ray Ancho

Load-bearing premise

The method assumes that a handful of samples along each camera ray, starting at the predicted visible surface, provides enough Gaussian anchors to cover the interior of every occupied object and the free space around it—if a solid object's interior does not lie on those sampled ray segments (thin, oblique, or textureless regions), that volume cannot be recovered.

What would settle it

Measure per-class IoU on Occ-ScanNet for thin and oblique structures such as chair legs, window frames, and open shelving: if the reported 57.4 mIoU is carried by the ray-sampling prior, these classes should show disproportionately large gaps against a dense-grid baseline. Alternatively, a controlled ablation that replaces ray-sampled anchors with randomly placed anchors (same Gaussian count and training budget) should show a large mIoU drop if the ray prior is the cause of the gains.

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

If this is right

  • Occupancy prediction can improve simply by swapping in stronger visual geometry models, since the ray-sampling transformation is agnostic to how the surface prior is produced.
  • The same sparse Gaussian representation covers monocular, streaming temporal, and multi-camera inputs, so a single framework can serve both embodied indoor perception and surround-view driving perception.
  • Concentrating Gaussians around surfaces and interiors rather than empty space keeps the representation compact, yielding faster inference than dense-grid baselines at comparable or better accuracy.
  • The reported gains on both indoor and outdoor benchmarks suggest the surface-to-volume transformation is a general recipe rather than a scene-type-specific fix.

Where Pith is reading between the lines

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

  • A natural extension the authors leave implicit is closing the loop: feeding predicted occupancy back as supervision or as auxiliary input to the geometry prior, which could improve surface estimates in precisely the occluded regions where ray sampling has no evidence.
  • Because the anchors live on camera rays, objects seen nearly edge-on or surfaces almost parallel to the ray may be under-covered; a per-class failure analysis on thin structures (chair legs, railings, window frames) would test whether this representational bias shows up as systematic mIoU gaps.
  • The incremental update is training-free, so it should work with on-the-fly poses estimated by the geometry model itself; whether the reported embodied performance depends on ground-truth camera poses is a testable question the paper does not fully separate.
  • The memory bank grows as new Gaussians are inserted over long sequences, so the efficiency claim may degrade on extended trajectories; measuring FPS and memory as a function of sequence length would quantify that trade-off.

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 extends the authors' prior GPOcc framework to GPOcc++, targeting camera-only 3D semantic occupancy prediction. The core idea is to convert surface-centric visual geometry priors (DepthAnything or VGGT) into sparse 3D Gaussian primitives through ray-based volumetric sampling, opacity pruning, and a probabilistic Gaussian-to-occupancy aggregation. The new technical contributions are Ray-Conditioned Multi-Image Fusion, which uses Plücker-ray embeddings and global attention over multiple views/temporal frames, and Offset-Guided Ray Anchoring, which predicts 3D residuals for the ray-sampled Gaussian centers. The method is evaluated on Occ-ScanNet, EmbodiedOcc-ScanNet, and nuScenes, with extensive ablations and efficiency profiling. The authors report state-of-the-art or competitive results, especially on indoor benchmarks, and release code.

Significance. If the reported results are reliable, GPOcc++ offers a meaningful step in unifying sparse Gaussian representations with visual geometry foundation models for both indoor embodied and outdoor driving occupancy prediction. The paper includes several concrete strengths: ablations isolating the contributions of each module (Tables IV–VI), efficiency comparisons (Tables VII–IX), and a public code release. The idea of ray-conditioned cross-view/temporal fusion and offset-guided anchoring is well motivated and the experimental coverage is broad. However, the headline claims are currently confounded by backbone and resolution mismatches, and some stated conclusions are not supported by the tables, so the significance cannot be fully assessed without additional controlled comparisons.

major comments (3)
  1. [§VI-C, Table I and Table V] The claim that GPOcc++ (VGGT) surpasses EmbodiedOcc++ by +11.21 mIoU is not a clean architecture comparison because the best variant uses a 942M-parameter frozen VGGT backbone whereas EmbodiedOcc++ uses DepthAnything. The paper's own ablations (Table V) show that switching the geometry prior from DPT to VGGT improves GPOcc by +4.31 mIoU and GPOcc++ by +3.53 mIoU, while the architecture-only gains are +2.00 (DPT) and +1.22 (VGGT). A same-backbone baseline (e.g., EmbodiedOcc++ with VGGT) is never reported, so the marginal contribution of the proposed sparse-Gaussian construction is not isolated. Please provide same-backbone, same-resolution comparisons or an explicit decomposition of the gains.
  2. [§VI-E, Table VI and §VI-C, Table III] The nuScenes ablation shows very small gains: Multi-Image Fusion adds +0.20 mIoU and the full GPOcc++ adds +0.95 mIoU over GPOcc, with no error bars or multiple seeds reported. Given that Table III lists QuadricFormer at 21.86 mIoU and GPOcc++ at 20.78 mIoU, the statement in §VI-F/Table IX that GPOcc++ achieves 'the best overall accuracy' is contradicted by these numbers. Please correct the claim and report variance or statistical significance for the small deltas.
  3. [§IV-B, Eq. (5)-(6) and §VII] The ray-based volumetric sampling assumes that sampling K linearly spaced offsets from each visible surface point along the camera ray is sufficient to cover occupied interiors and free space. Table IV shows this is essential (mIoU drops from 56.19 to 47.88 without it), and the Limitations section concedes weak coverage on large flat/textureless regions. The manuscript should provide a more detailed analysis of failure cases (oblique surfaces, thin structures) or an additional mechanism to ensure coverage, since the core representation hinges on this geometric postulate.
minor comments (5)
  1. [Abstract and §I] The abstract introduces 'GPOcc' as the main contribution, but the title and the rest of the paper focus on GPOcc++. Please harmonize the wording and explicitly state that this is an extension of the conference paper [28].
  2. [Table I] Formatting issue: the entry '52.0746.44' in the GPOcc (VGGT) row should be '52.07 46.44'. Please proofread all tables for missing spaces and alignment.
  3. [§VI-B and Table IX] The FPS comparisons on nuScenes use 518px input while compared methods use 900×1600 resolution. This is acknowledged in the text, but the tables should include a footnote so the reader does not infer a fair efficiency comparison at a glance.
  4. [§V-B] The causal mask for temporal observations is mentioned, but it is not clear whether the mask is applied per frame or across the concatenated token sequence. Please clarify in the text or figure.
  5. [References] Reference [51] is dated 2026 with an arXiv identifier '2604.28115', which appears inconsistent with the current timeline. Please verify the citation accuracy.

Circularity Check

0 steps flagged

No significant circularity; the core claims are validated on external benchmarks and the only self-citation is a normal, acknowledged extension of the authors' own prior GPOcc.

full rationale

The paper is an empirical architecture paper. Its central transformation, ray-based volumetric sampling (Eqs. 5-6), is a modeling choice trained against ground-truth occupancy on Occ-ScanNet, EmbodiedOcc-ScanNet, and nuScenes; it is not a parameter fitted to a subset and then reported as a prediction of the same quantity. The Gaussian-to-occupancy aggregation follows GaussianFormer-2 [18], an external source, and the depth/geometry priors (DepthAnything, VGGT) are external foundation models. The only notable self-citation is [28], the authors' own CVPR GPOcc paper, which the article explicitly builds upon and acknowledges. This is a normal research-extension chain: GPOcc++ is compared against GPOcc and against external methods, with ablations isolating the two new modules (Tables V-VI). No load-bearing conclusion is justified solely by an unverified self-citation, and no 'uniqueness theorem' or ansatz is imported from the authors' prior work. The limitations discussion (Section VII) concedes weakness on large flat/textureless regions; this is an accuracy limitation of the geometric ray-sampling assumption, not a circularity. The skeptic's concerns about backbone mismatch and lack of same-backbone controls are evaluation/attribution risks, not derivation-circularity, and therefore do not raise the circularity score above the minor-self-citation band.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method depends on frozen visual-geometry foundation models, a geometric postulate about ray-sampling coverage, the Gaussian super-position occupancy model from GaussianFormer-2, known camera poses, and benchmark annotations. No new physical entities are introduced; the sparse Gaussians and learnable embeddings are trainable parameters, not independent postulates.

free parameters (5)
  • K (ray samples per pixel) = 16
    Number of volumetric anchors per pixel in Eq. (5)-(6); set by hand with sensitivity analysis deferred to appendix; directly controls representation capacity and compute.
  • tau (opacity pruning threshold) = 0.01
    Primitives with opacity below tau are discarded; chosen by default, affects the compactness vs recall trade-off (Table IV).
  • gamma (incremental update balance) = <0.5 (unspecified exact value)
    Balance between historical and new Gaussian attributes in Eq. (10); the paper only states gamma<0.5 to favor recent observations, leaving the exact value under-specified.
  • epsilon (neighbor radius for incremental fusion) = not reported
    Spatial radius used to find matching memory Gaussians for confidence-weighted averaging in Eq. (10); never quantified in the text.
  • loss weights for Eq. (11) = not reported
    Composite loss sums focal, Lovász, two affinity, and Huber depth terms without stated weights, affecting training behavior and the relative importance of geometry vs occupancy supervision.
axioms (5)
  • domain assumption A frozen visual geometry prior (VGGT or DepthAnything-V2) provides sufficiently accurate surface geometry and features for anchor generation.
    The entire pipeline in Secs. IV-B and V uses VGGT/DPT outputs (depth, pointmaps, features) as the geometric substrate; in textureless/flat regions the paper's own Limitations section notes the priors give weak evidence.
  • ad hoc to paper Sampling K points along the camera ray from the surface into the volume is sufficient to cover occupied interiors and free space (Eqs. 5-6).
    No theoretical guarantee that object interiors lie on the sampled ray segment; ablation (Table IV) shows removing ray sampling hurts by 8.31 mIoU, so the central mechanism rests on this geometric postulate.
  • domain assumption Occupancy can be obtained by weighted superposition of Gaussian kernels (Eq. 1).
    Adopted from GaussianFormer-2 [18] without independent derivation; the choice of Gaussian kernel shape and the neighbor set N(p) influence the mapping.
  • domain assumption Accurate camera poses are available for fusion and incremental update.
    Eqs. (10) and (13)-(16) require transforming Gaussians and rays into a global coordinate system; the method assumes poses are given rather than estimated.
  • domain assumption Ground-truth occupancy annotations (SurroundOcc nuScenes, Occ-ScanNet) are reliable and the eval protocol matches prior work.
    All numbers in Tables I-III are measured against these labels; the SurroundOcc [13] annotations are themselves pseudo-labels generated by another model, which is not discussed.

pith-pipeline@v1.3.0-alltime-deepseek · 20479 in / 16559 out tokens · 164137 ms · 2026-08-02T05:02:29.194718+00:00 · methodology

0 comments
read the original abstract

Accurate 3D scene understanding is fundamental to embodied intelligence and autonomous driving, where 3D occupancy provides a unified representation of objects, structures, and free space. However, recovering such a complete volumetric representation from visual observations remains challenging, particularly in occluded and unobserved regions. Visual geometry priors offer strong and generalizable geometric cues for addressing this challenge, but their outputs are inherently surface-centric, whereas occupancy prediction requires reasoning about volumetric interiors and free space. To bridge this gap, we introduce GPOcc, which transforms visual geometry priors into occupancy-aware sparse Gaussian representations for efficient and expressive volumetric scene modeling. Building on GPOcc, GPOcc++ models multi-view observations and temporal sequences within a unified framework, allowing spatial and temporal evidence to be handled through the same representation. We further extend GPOcc++ from indoor scenes to outdoor occupancy prediction. Extensive experiments on both indoor and outdoor benchmarks demonstrate consistently strong performance across both multi-view and temporal settings, together with favorable efficiency and generalization. Code will be released at https://github.com/JuIvyy/GPOcc.

Figures

Figures reproduced from arXiv: 2607.13481 by Bing Wang, Changhao Chen, Changqing Zhou, Jie Qin, Yueru Luo, Yulan Guo.

Figure 1
Figure 1. Figure 1: Comparison of monocular occupancy prediction pipelines. ISO [1] formulates depth estimation as a multi-class classification problem, using the predicted depth distributions to lift 2D image features into dense 3D volumes, which are then processed by a 3D U-Net for occupancy prediction. EmbodiedOcc [2], by contrast, initializes random 3D anchors and applies cross-attention to aggregate image features, predi… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of GPOcc. Given an RGB image, a visual geometry prior predicts surface geometry and extracts 3D￾aware features. Ray-based volumetric sampling extends the predicted surface points into interior volumetric anchors, which are represented by sparse semantic Gaussian primitives. The resulting Gaussians are probabilistically mapped to local occupancy. For streaming observations, per-frame Gaussian predi… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of Gaussian representations. (a) Em￾bodiedOcc places many Gaussian primitives in empty regions, shown in gray. (b) GPOcc concentrates sparse Gaussians on and within occupied structures, resulting in a more compact representation. C. From Sparse Gaussians to Occupancy Prior Gaussian occupancy methods [2] initialize a dense set of predefined 3D anchors and classify them as occupied or empty. Since… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of GPOcc++. Given temporal or multi-view images, a visual geometry model extracts per-image features and surface predictions, while Plucker embeddings encode the corresponding camera rays. Ray-Conditioned Multi-Image Fusion ¨ models cross-view or cross-temporal relationships in a unified feature space. The surface predictions define initial ray-based volumetric anchors, and the fused features cond… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on monocular occupancy prediction. (a) shows the input RGB images, (b) the ground-truth occupancy, (c) the predictions of EmbodiedOcc [2], (d) the predictions of GPOcc, (e) the predictions of GPOcc++, and (f) the Gaussian primitives predicted by GPOcc++. Compared to EmbodiedOcc and GPOcc, GPOcc++ produces more accurate and complete reconstructions, while the Gaussian representation p… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results on streaming inputs. (a) shows the input RGB images, (b) the ground-truth occupancy, (c) the predictions of GPOcc, and (d) the predictions of GPOcc++. Our incremental update strategy progressively integrates information from sequential frames, and GPOcc++ produces more complete and consistent predictions. threshold τ , together with ablations on the incremental update strategy, are prov… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results on nuScenes. We visualize surround-view camera inputs, the semantic occupancy predicted by GPOcc++, and the corresponding ground-truth occupancy in outdoor driving scenes. TABLE VI: GPOcc++ component ablation on nuScenes. Setting IoU mIoU GPOcc 31.92 19.83 + Multi-Image Fusion 32.04 20.03 GPOcc++ 33.28 20.78 Effect of GPOcc++ components on nuScenes. Table VI shows the corresponding GPOc… 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

65 extracted references · 11 linked inside Pith

  1. [1]

    Monocular occupancy prediction for scalable indoor scenes,

    H. Yu, Y . Wang, Y . Chen, and Z. Zhang, “Monocular occupancy prediction for scalable indoor scenes,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 38–54

    2024

  2. [2]

    Embodiedocc: Embodied 3d occupancy prediction for vision-based online scene under- standing,

    Y . Wu, W. Zheng, S. Zuo, Y . Huang, J. Zhou, and J. Lu, “Embodiedocc: Embodied 3d occupancy prediction for vision-based online scene under- standing,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 26 360–26 370

    2025

  3. [3]

    Embodied intelligence: A synergy of morphology, action, perception and learning,

    H. Liu, D. Guo, and A. Cangelosi, “Embodied intelligence: A synergy of morphology, action, perception and learning,”ACM Computing Surveys, vol. 57, no. 7, pp. 1–36, 2025

    2025

  4. [4]

    Bevformer: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers,

    Z. Li, W. Wang, H. Li, E. Xie, C. Sima, T. Lu, Q. Yu, and J. Dai, “Bevformer: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  5. [5]

    Unimode: Unified monocular 3d object detection,

    Z. Li, X. Xu, S. Lim, and H. Zhao, “Unimode: Unified monocular 3d object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 561–16 570

    2024

  6. [6]

    Toward practical monocular indoor depth estimation,

    C.-Y . Wu, J. Wang, M. Hall, U. Neumann, and S. Su, “Toward practical monocular indoor depth estimation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 3814– 3824

    2022

  7. [7]

    Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection,

    D. Rukhovich, A. V orontsova, and A. Konushin, “Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2022, pp. 2397–2406

    2022

  8. [8]

    Learning occupancy for monocular 3d object detection,

    L. Peng, J. Xu, H. Cheng, Z. Yang, X. Wu, W. Qian, W. Wang, B. Wu, and D. Cai, “Learning occupancy for monocular 3d object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10 281–10 292

    2024

  9. [9]

    V olumetric environment representation for vision-language navigation,

    R. Liu, W. Wang, and Y . Yang, “V olumetric environment representation for vision-language navigation,” inProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, 2024, pp. 16 317– 16 328

    2024

  10. [10]

    Occvla: Vision-language- action model with implicit 3d occupancy supervision,

    R. Liu, L. Kong, D. Li, and H. Zhao, “Occvla: Vision-language- action model with implicit 3d occupancy supervision,”arXiv preprint arXiv:2509.05578, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Occllama: An occupancy-language-action generative world model for autonomous driving,

    J. Wei, S. Yuan, P. Li, Q. Hu, Z. Gan, and W. Ding, “Occllama: An occupancy-language-action generative world model for autonomous driving,”arXiv preprint arXiv:2409.03272, 2024. 13

    Pith/arXiv arXiv 2024

  12. [12]

    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,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 9087– 9098

    2023

  13. [13]

    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 Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 21 729–21 740

    2023

  14. [14]

    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,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 9223–9232

    2023

  15. [15]

    Occupancy as set of points,

    Y . Shi, T. Cheng, Q. Zhang, W. Liu, and X. Wang, “Occupancy as set of points,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 72–87

    2024

  16. [16]

    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,”Advances in Neural Information Processing Systems, vol. 37, pp. 119 861–119 885, 2024

    2024

  17. [17]

    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,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 376– 393

    2024

  18. [18]

    Gaussianformer-2: Probabilistic gaussian superposition for effi- cient 3d occupancy prediction,

    Y . Huang, A. Thammatadatrakoon, W. Zheng, Y . Zhang, D. Du, and J. Lu, “Gaussianformer-2: Probabilistic gaussian superposition for effi- cient 3d occupancy prediction,” inProceedings of the computer vision and pattern recognition conference, 2025, pp. 27 477–27 486

    2025

  19. [19]

    Depth anything: Unleashing the power of large-scale unlabeled data,

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth anything: Unleashing the power of large-scale unlabeled data,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 10 371–10 381

    2024

  20. [20]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,”Advances in Neural Information Processing Sys- tems, vol. 37, pp. 21 875–21 911, 2024

    2024

  21. [21]

    Bevheight++: Toward robust visual centric 3d object detection,

    L. Yang, T. Tang, J. Li, K. Yuan, K. Wu, P. Chen, L. Wang, Y . Huang, L. Li, X. Zhanget al., “Bevheight++: Toward robust visual centric 3d object detection,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  22. [22]

    Vggt: Visual geometry grounded transformer,

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “Vggt: Visual geometry grounded transformer,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5294– 5306

    2025

  23. [23]

    Dust3r: Geometric 3d vision made easy,

    S. Wang, V . Leroy, Y . Cabon, B. Chidlovskii, and J. Revaud, “Dust3r: Geometric 3d vision made easy,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 697–20 709

    2024

  24. [24]

    Grounding image matching in 3d with mast3r,

    V . Leroy, Y . Cabon, and J. Revaud, “Grounding image matching in 3d with mast3r,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 71–91

    2024

  25. [25]

    Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass,

    J. Yang, A. Sax, K. J. Liang, M. Henaff, H. Tang, A. Cao, J. Chai, F. Meier, and M. Feiszli, “Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 21 924–21 935

    2025

  26. [26]

    π 3: Scalable permutation-equivariant visual geometry learning,

    Y . Wang, J. Zhou, H. Zhu, W. Chang, Y . Zhou, Z. Li, J. Chen, J. Pang, C. Shen, and T. He, “π 3: Scalable permutation-equivariant visual geometry learning,”arXiv preprint arXiv:2507.13347, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Point3r: Streaming 3d recon- struction with explicit spatial pointer memory,

    Y . Wu, W. Zheng, J. Zhou, and J. Lu, “Point3r: Streaming 3d recon- struction with explicit spatial pointer memory,”Advances in Neural Information Processing Systems, vol. 38, pp. 69 675–69 699, 2026

    2026

  28. [28]

    Generalizing visual geometry priors to sparse gaussian occupancy prediction,

    C. Zhou, Y . Luo, and C. Chen, “Generalizing visual geometry priors to sparse gaussian occupancy prediction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 28 578–28 587

    2026

  29. [29]

    Towards robust monocular depth estimation: Mixing datasets for zero-shot cross- dataset transfer,

    R. Ranftl, K. Lasinger, D. Hafner, K. Schindler, and V . Koltun, “Towards robust monocular depth estimation: Mixing datasets for zero-shot cross- dataset transfer,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 3, pp. 1623–1637, 2020

    2020

  30. [30]

    Raymvsnet++: learning ray-based 1d implicit fields for accurate multi-view stereo,

    Y . Shi, J. Xi, D. Hu, Z. Cai, and K. Xu, “Raymvsnet++: learning ray-based 1d implicit fields for accurate multi-view stereo,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 11, pp. 13 666–13 682, 2023

    2023

  31. [31]

    3d reconstruction with spatial memory,

    H. Wang and L. Agapito, “3d reconstruction with spatial memory,” in 2025 International Conference on 3D Vision (3DV). IEEE, 2025, pp. 78–89

    2025

  32. [32]

    Con- tinuous 3d perception model with persistent state,

    Q. Wang, Y . Zhang, A. Holynski, A. A. Efros, and A. Kanazawa, “Con- tinuous 3d perception model with persistent state,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 10 510–10 522

    2025

  33. [33]

    Must3r: Multi-view network for stereo 3d reconstruction,

    Y . Cabon, L. Stoffl, L. Antsfeld, G. Csurka, B. Chidlovskii, J. Revaud, and V . Leroy, “Must3r: Multi-view network for stereo 3d reconstruction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 1050–1060

    2025

  34. [34]

    Streaming 4d visual geometry transformer,

    D. Zhuo, W. Zheng, J. Guo, Y . Wu, J. Zhou, and J. Lu, “Streaming 4d visual geometry transformer,”arXiv preprint arXiv:2507.11539, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Dens3r: A foundation model for 3d geometry prediction,

    X. Fang, J. Gao, Z. Wang, Z. Chen, X. Ren, J. Lyu, Q. Ren, Z. Yang, X. Yang, Y . Yanet al., “Dens3r: A foundation model for 3d geometry prediction,”arXiv preprint arXiv:2507.16290, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Monoscene: Monocular 3d semantic scene completion,

    A.-Q. Cao and R. De Charette, “Monoscene: Monocular 3d semantic scene completion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 3991–4001

    2022

  37. [37]

    Ndc- scene: Boost monocular 3d semantic scene completion in normalized device coordinates space,

    J. Yao, C. Li, K. Sun, Y . Cai, H. Li, W. Ouyang, and H. Li, “Ndc- scene: Boost monocular 3d semantic scene completion in normalized device coordinates space,” in2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE Computer Society, 2023, pp. 9421– 9431

    2023

  38. [38]

    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,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1746–1754

    2017

  39. [39]

    Anisotropic convolutional neural networks for rgb-d based semantic scene completion,

    J. Li, P. Wang, K. Han, and Y . Liu, “Anisotropic convolutional neural networks for rgb-d based semantic scene completion,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 8125–8138, 2021

    2021

  40. [40]

    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 transactions on pattern analysis and machine intelligence, vol. 44, no. 10, pp. 7205–7218, 2021

    2021

  41. [41]

    Ssr-2d: semantic 3d scene reconstruction from 2d images,

    J. Huang, A. Artemov, Y . Chen, S. Zhi, K. Xu, and M. Nießner, “Ssr-2d: semantic 3d scene reconstruction from 2d images,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 8486– 8501, 2024

    2024

  42. [42]

    Vision-centric bev perception: A survey,

    Y . Ma, T. Wang, X. Bai, H. Yang, Y . Hou, Y . Wang, Y . Qiao, R. Yang, and X. Zhu, “Vision-centric bev perception: A survey,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 10 978–10 997, 2024

    2024

  43. [43]

    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,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2024, pp. 15 035– 15 044

    2024

  44. [44]

    Delving into the devils of bird’s-eye-view perception: A review, evaluation and recipe,

    H. Li, C. Sima, J. Dai, W. Wang, L. Lu, H. Wang, J. Zeng, Z. Li, J. Yang, H. Denget al., “Delving into the devils of bird’s-eye-view perception: A review, evaluation and recipe,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 4, pp. 2151–2170, 2023

    2023

  45. [45]

    Occscene: Semantic occupancy-based cross-task mutual learning for 3d scene generation,

    B. Li, X. Jin, J. Wang, Y . Shi, Y . Sun, X. Wang, Z. Ma, B. Xie, C. Ma, X. Yanget al., “Occscene: Semantic occupancy-based cross-task mutual learning for 3d scene generation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  46. [46]

    Spot: Scalable 3d pre-training via occupancy prediction for learning transferable 3d representations,

    X. Yan, R. Chen, B. Zhang, H. Ye, R. Xia, J. Yuan, H. Zhou, X. Cai, B. Shi, W. Shaoet al., “Spot: Scalable 3d pre-training via occupancy prediction for learning transferable 3d representations,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  47. [47]

    Omnihd-scenes: A next-generation multimodal dataset for autonomous driving,

    L. Zheng, L. Yang, Q. Lin, W. Ai, M. Liu, S. Lu, J. Liu, H. Ren, J. Mo, X. Baiet al., “Omnihd-scenes: A next-generation multimodal dataset for autonomous driving,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  48. [48]

    Hybrid- prediction integrated planning for autonomous driving,

    H. Liu, Z. Huang, W. Huang, H. Yang, X. Mo, and C. Lv, “Hybrid- prediction integrated planning for autonomous driving,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 47, no. 4, pp. 2597–2614, 2025

    2025

  49. [49]

    Embodiedocc++: Boosting embodied 3d occupancy prediction with plane regularization and uncertainty sampler,

    H. Wang, X. Wei, X. Zhang, J. Li, C. Bai, Y . Li, M. Lu, W. Zheng, and S. Zhang, “Embodiedocc++: Boosting embodied 3d occupancy prediction with plane regularization and uncertainty sampler,”arXiv preprint arXiv:2504.09540, 2025

    Pith/arXiv arXiv 2025

  50. [50]

    Monocular open vocabulary occupancy prediction for indoor scenes,

    C. Zhou, Y . Luo, H. Zhang, Z. Jiang, and C. Chen, “Monocular open vocabulary occupancy prediction for indoor scenes,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 21 627–21 637

    2026

  51. [51]

    Freeocc: Training- free embodied open-vocabulary occupancy prediction,

    Z. Jiang, C. Zhou, X. Zuo, and C. Chen, “Freeocc: Training- free embodied open-vocabulary occupancy prediction,”arXiv preprint arXiv:2604.28115, 2026. 14

    Pith/arXiv arXiv 2026

  52. [52]

    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

  53. [53]

    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,” inEuropean conference on computer vision. Springer, 2020, pp. 194–210

    2020

  54. [54]

    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,”Advances in Neural Information Processing Systems, vol. 37, pp. 79 618–79 641, 2024

    2024

  55. [55]

    Visibility-aware point-based multi-view stereo network,

    R. Chen, S. Han, J. Xu, and H. Su, “Visibility-aware point-based multi-view stereo network,”IEEE transactions on pattern analysis and machine intelligence, vol. 43, no. 10, pp. 3695–3708, 2020

    2020

  56. [56]

    Visibility-aware multi-view stereo by surface normal weighting for occlusion robustness,

    H. Lee, S. Lee, and S. Lee, “Visibility-aware multi-view stereo by surface normal weighting for occlusion robustness,”IEEE transactions on pattern analysis and machine intelligence, 2025

    2025

  57. [57]

    Fast transformer decoding: One write-head is all you need,

    N. Shazeer, “Fast transformer decoding: One write-head is all you need,” arXiv preprint arXiv:1911.02150, 2019

    Pith/arXiv arXiv 1911

  58. [58]

    Xvii. on a new geometry of space,

    J. Plucker, “Xvii. on a new geometry of space,”Philosophical Transac- tions of the Royal Society of London, no. 155, pp. 725–791, 1865

  59. [59]

    Rayzer: A self-supervised large view synthesis model,

    H. Jiang, H. Tan, P. Wang, H. Jin, Y . Zhao, S. Bi, K. Zhang, F. Luan, K. Sunkavalli, Q. Huanget al., “Rayzer: A self-supervised large view synthesis model,” inProceedings of the IEEE/CVF International Con- ference on Computer Vision, 2025, pp. 4918–4929

    2025

  60. [60]

    Roboocc: Enhancing the geometric and semantic scene understanding for robots,

    Z. Zhang, Q. Zhang, W. Cui, S. Shi, Y . Guo, G. Han, W. Zhao, H. Ren, R. Xu, and J. Tang, “Roboocc: Enhancing the geometric and semantic scene understanding for robots,”arXiv preprint arXiv:2504.14604, 2025

    Pith/arXiv arXiv 2025

  61. [61]

    Atlas: End-to-end 3d scene reconstruction from posed images,

    Z. Murez, T. v. As, J. Bartolozzi, A. Sinha, V . Badrinarayanan, and A. Rabinovich, “Atlas: End-to-end 3d scene reconstruction from posed images,” inECCV, 2020, pp. 414–431

    2020

  62. [62]

    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,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp. 9433–9443

    2023

  63. [63]

    Quadricformer: Scene as superquadrics for 3d semantic occupancy prediction,

    S. Zuo, W. Zheng, X. Han, L. Yang, J. Luet al., “Quadricformer: Scene as superquadrics for 3d semantic occupancy prediction,”Advances in Neural Information Processing Systems, vol. 38, pp. 47 779–47 801, 2026

    2026

  64. [64]

    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,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 621–11 631

    2020

  65. [65]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017. Changqing Zhouis currently pursuing the Ph.D. degree at The Hong Kong University of Science and Technology (Guangzhou). His research inter- ests include 3D computer vision, embodied scene understanding, occupancy prediction, and open- vocabulary 3D...

    Pith/arXiv arXiv 2017