Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

Long-horizon indoor occupancy maps work better when local geometry is treated as transient evidence, not the map itself.

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 06:03 UTC pith:C2NOLGQT

load-bearing objection Useful hierarchical indoor occupancy benchmark plus a coherent evidence-vs-memory design; local gains over GPOcc are small, but online free-space and building-scale results are the real story. the 4 major comments →

arxiv 2607.05543 v1 pith:C2NOLGQT submitted 2026-07-06 cs.RO cs.CV

GEM-Occ: From Visual Geometry Evidence to Embodied Semantic Occupancy Memory

Hu Zhu , Bohan Li , Xianda Guo , Hongsi Liu , Baorui Peng , Mingqi Yuan , Xin Jin , Wenjun Zeng
show 1 more author
Chang Wen Chen
This is my paper
classification cs.RO cs.CV
keywords semantic occupancyembodied mappingGaussian evidence memoryhierarchical indoor occupancyfree-space reasoningonline 3D mappingHIOccrobotics perception
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.

Embodied agents need a persistent 3D memory of occupied structure, free space, unknown regions, and object labels as they move through rooms and buildings. Most indoor occupancy work still stops at single views or one room, and pointmaps or per-frame predictions do not make a stable map when observations arrive over time. This paper builds HIOcc so local, room-level, and building-level mapping can be measured under one sparse occupancy format across perspective and panoramic indoor data. It then argues that the right map state is a hierarchical Gaussian evidence memory: each view is turned into semantic occupancy Gaussians and free-space ray evidence, fused with visibility- and uncertainty-aware updates, and read out by Gaussian-to-occupancy splatting. On HIOcc, that design improves local accuracy, online stability, free-space reasoning, revisit consistency, and building-scale efficiency over prior indoor occupancy and Gaussian mapping baselines.

Core claim

Pointmaps and other local visual geometry predictions should be treated as transient evidence, not as persistent map state. Converting them into semantic Gaussian occupancy evidence plus free-space ray evidence, then fusing those signals with visibility- and uncertainty-aware causal updates into a hierarchical memory of local caches, room submaps, and a building graph, yields more accurate, stable, and scalable semantic occupancy maps for embodied indoor agents than voxel lifting, pointmap fusion, or frame-level Gaussian prediction.

What carries the argument

GEM-Occ: a Gaussian Evidence Memory that turns each posed observation into world-frame semantic Gaussians and free-space rays, matches and fuses them by confidence-weighted, visibility-aware causal updates, and queries the resulting hierarchical map by Gaussian-to-occupancy splatting.

Load-bearing premise

Local depth or pointmap estimates from the view encoder are reliable enough that confidence-weighted Gaussian fusion will not lock surface errors into the map near edges, transparent surfaces, or weak texture.

What would settle it

On the same HIOcc building-level streams, replace the local geometry prior with one that is deliberately noisy near boundaries and weak texture; if free-space and revisit metrics no longer beat flat Gaussian or pointmap fusion, the evidence-to-memory claim fails.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Indoor occupancy evaluation should report local prediction, causal room mapping, and building-scale mapping together rather than single-room scores alone.
  • Map systems should keep observed free space separate from unknown space instead of filling unobserved volumes as empty.
  • Hierarchical Gaussian memory can cut per-area storage and query cost while improving revisit consistency over a flat global Gaussian store.
  • Semantic Gaussians are usable as long-lived map elements for streaming agents, not only as per-frame prediction primitives.

Where Pith is reading between the lines

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

  • If free-space ray evidence is the main driver of online gains, active exploration policies that deliberately cast rays through uncertain free corridors should improve map quality faster than pure surface following.
  • The same evidence-vs-memory split may transfer to outdoor or mixed indoor-outdoor agents once panoramic and multi-floor connectivity are replaced by route graphs.
  • Downstream navigation and object-search agents would be a sharper test of whether hierarchical occupancy memory actually changes behavior, not only map metrics.

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

4 major / 7 minor

Summary. The paper introduces HIOcc, a hierarchical indoor semantic occupancy benchmark that unifies ScanNet, ScanNet++, and Matterport3D under a common sparse occupancy format while preserving native perspective and panoramic observation geometries, with evaluation protocols for local prediction, room-level online mapping, and building-level panoramic mapping. It further proposes GEM-Occ, which treats local visual geometry (depth/pointmaps) as transient evidence, converts it into semantic Gaussian occupancy primitives and free-space ray evidence, and fuses these into a hierarchical persistent memory (local caches, room submaps, building graph) via visibility- and uncertainty-aware causal updates, with Gaussian-to-occupancy splatting for queries. On HIOcc, GEM-Occ reports gains over voxel, pointmap-fusion, and Gaussian occupancy baselines on local IoU/mIoU, room-level online mapping, revisit consistency, free-space reasoning, and building-level accuracy/scalability metrics.

Significance. If the results hold under tighter statistical and baseline controls, the work is a useful systems contribution for embodied indoor perception. HIOcc fills a clear gap by jointly supporting perspective and panoramic observations and local-to-building hierarchical mapping under one sparse occupancy protocol; the annotation pipeline and visibility-aware validation (Tables 2, 6) are carefully documented. The evidence-versus-memory framing—semantic Gaussians as long-lived map elements plus an explicit free-space ray cache—is a coherent design response to surface-biased pointmap accumulation, and the hierarchical memory results (Table 4) address scalability in a way prior room-bounded occupancy methods largely do not. Strengths include multi-regime evaluation, component ablations, and explicit free/unknown separation. The contribution is empirical and engineering-oriented rather than theoretical; its lasting value depends on HIOcc adoption and on whether online/building gains are shown to come from the fusion design rather than mainly from the local geometry encoder.

major comments (4)
  1. Table 3 local prediction: GEM-Occ’s occupancy IoU over the strongest Gaussian baseline GPOcc is only +0.68 (61.37 vs 60.69), while the abstract packages local prediction, online stability, free-space reasoning, revisit consistency, and building scalability as jointly improved. No error bars, multi-seed variance, or significance tests are reported for any table. Without that, it is unclear whether the local margin is reliable, and whether room/building gains (Tables 3–4) partly reflect a stronger local encoder (Dt, St, Qt in §4 / C.1) rather than the evidence-vs-memory design. Please report variance (or multi-seed runs) for local and room metrics and, if possible, a controlled comparison that freezes the same local geometry backbone across fusion variants.
  2. Table 5 vs Appendix Table 9: the main-paper ablation uses an inconsistent semantic mIoU definition that includes free space, while the corrected 11-class Table 9 changes several entries (e.g., w/o hierarchical memory local mIoU 52.41→50.29; free-space ablation local mIoU 51.36→52.36; pointmap room mIoU 44.72→41.17). The paper itself flags this in D.3. Main-text claims about component importance currently rest on the incorrect table. Replace Table 5 with the corrected protocol, restate the ablation conclusions from Table 9 only, and ensure all main-text mIoU numbers use the same 11-class definition as Tables 3, 7, and 8.
  3. §4 / C.1–C.2 and Limitations §6: the method unprojects local geometry Dt into world-frame Gaussians and fuses them causally; Limitations correctly note error propagation near boundaries, transparent surfaces, and weak texture. This is load-bearing for the online and building claims, which reuse the same local evidence. The manuscript does not quantify how geometry errors affect memory (e.g., sensitivity to depth noise, confidence calibration of η, or failure cases on glass/mirrors). Add at least a targeted stress analysis or qualitative failure breakdown so the free-space ray and confidence-weighted fusion claims can be assessed under realistic prior error.
  4. Table 4 building-level baselines (per-frame fusion, pointmap fusion, post-memory fusion, flat Gaussian memory) are weaker and less aligned with the EmbodiedOcc / GPOcc line used in Table 3. The hierarchical-memory win (53.8 IoU, 0.81 MB/m² vs flat 50.4 / 1.36) is informative, but without stronger published online occupancy baselines under the same Matterport3D HIOcc protocol, the building-level superiority claim is only partially secured. Either port EmbodiedOcc/GPOcc-style memory methods to this split or clearly state that Table 4 is an internal design comparison, not a full external SOTA ranking.
minor comments (7)
  1. §4 and C.2: free parameters τ_m, τ_occ, λ, Δℓ_occ, Δℓ_free, τ_sem, and loss weights are named but numerical values and selection procedure are not given in the main text or appendix excerpts. Please list them for reproducibility.
  2. Eq. (1) / (13): the product-form occupancy accumulation is standard but not compared to additive log-odds or max pooling; a one-sentence justification or small readout ablation would help.
  3. Figure 4 qualitative comparison is useful but hard to read at manuscript scale; consider larger crops or per-class error maps for failure modes (windows, thin furniture).
  4. Table 1 coverage area units mix km² (outdoor) and m² (indoor); clarify that HIOcc’s ~142.6K m² is scene footprint, not route length, to avoid miscomparison with driving benchmarks.
  5. Typos / wording: “Occpancy” in Figure 2; “Vo x e l i z a t i o n” spacing artifacts in figures; “Completedenotes” missing space in Table 2 caption; consistent hyphenation of “room-level” / “building-level”.
  6. §5 Setup: state whether poses are ground-truth or estimated; online mapping claims depend on this, and SLAM noise would stress the Mahalanobis match gate.
  7. D.4 single-frame FPS (8.7) is reported without hardware details in the main experimental setup; move a short runtime/hardware note into §5 for context with Table 4 query latency.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper with external occupancy metrics and design-choice fusion, not definitional or fit-forced predictions.

full rationale

GEM-Occ’s load-bearing claims are empirical: occupancy IoU/mIoU, progress AUC, revisit consistency, memory, and latency on HIOcc splits against external baselines (Tables 3–5, 7–9). Local geometry (Dt, pointmaps) is treated as transient evidence converted into Gaussians and free-space rays, then fused by explicit causal rules (λ-weighted means, log-odds updates, free-space penalties, hierarchical submaps) and queried by Gaussian-to-occupancy splatting—none of which algebraically restates the reported scores. Training losses (Locc, Lsem, Lgeo, Lray) supervise against view-conditioned HIOcc targets; evaluation is held-out comparison, not a fitted parameter renamed as prediction. Self-citations (EmbodiedOcc, GPOcc, related Gaussian occupancy work) supply background baselines and motivation, not uniqueness theorems or ansatzes that force the result. Ablations remove components and measure degradation on the same external metrics, which is independent content rather than circular reduction. No self-definitional loop, fitted-input-as-prediction, or load-bearing self-citation chain is present.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The paper’s claims rest on standard 3D vision math, domain assumptions about static indoor scenes and usable visual geometry, and several hand-chosen fusion/query thresholds. Invented entities are the benchmark and the memory framework itself; they are engineering constructs with experimental handles, not new physical objects.

free parameters (6)
  • Mahalanobis match threshold τ_m
    Controls whether incoming Gaussian evidence merges or inserts; chosen for fusion behavior, not derived.
  • Occupancy query threshold τ_occ
    Converts continuous Pocc into occupied labels for evaluation.
  • Temporal decay λ and log-odds increments Δℓ_occ / Δℓ_free
    Hand-set causal fusion strengths for occupied and free-space evidence.
  • Semantic merge threshold τ_sem (symmetric KL)
    Governs pruning/merging of memory Gaussians for long-horizon efficiency.
  • Loss weights λ_occ, λ_sem, λ_geo, λ_ray
    Training trade-offs among occupancy, semantics, geometry, and free-space ray losses.
  • Voxel grid resolutions and crop sizes (e.g., 0.02 m / 240×240×144; 0.05 m / 120×120×60)
    Benchmark design choices that define targets and reported IoU.
axioms (5)
  • domain assumption Indoor scenes are static during mapping; moving objects and layout changes are out of scope.
    Stated in Limitations §6; fusion treats evidence as accumulating fixed structure.
  • domain assumption Local visual geometry (depth/pointmaps) plus semantics can be treated as transient but informative evidence after confidence weighting.
    Core of Method §4 / C.1; errors may still propagate near hard surfaces.
  • domain assumption Observed free space is the ray segment before the first surface hit; space behind remains unknown until observed.
    Visibility model used for free-space ray evidence and negative log-odds updates.
  • ad hoc to paper Gaussian-to-occupancy probabilistic accumulation (product form over nearby primitives) is a valid readout of occupancy probability.
    Eq. (1)/(13) design choice for querying memory; standard-ish but not uniquely forced.
  • standard math Standard SE(3) camera geometry, unprojection, and anisotropic ray-aligned covariances.
    Used throughout evidence construction.
invented entities (2)
  • HIOcc hierarchical indoor occupancy benchmark independent evidence
    purpose: Unified sparse semantic occupancy targets and protocols for local, room, and building evaluation across three source datasets.
    New dataset/protocol construct; quality checked via 2D–3D consistency ablations, not independently pre-existing.
  • GEM-Occ hierarchical Gaussian Evidence Memory (semantic Gaussians + free-space ray cache + local/room/building graph) independent evidence
    purpose: Persistent map state that is not a pointmap or dense global voxel tensor.
    Framework introduced by the paper; evaluated via ablations and baselines, but the entity is defined by the method.

pith-pipeline@v1.1.0-grok45 · 23165 in / 3432 out tokens · 30578 ms · 2026-07-11T06:03:08.998573+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of GEM-Occ: From Visual Geometry Evidence to Embodied Semantic Occupancy Memory." pith.science (2026). https://pith.science/paper/C2NOLGQT

@misc{pith2026260705543,
  author       = {Pith},
  title        = {Pith review of: GEM-Occ: From Visual Geometry Evidence to Embodied Semantic Occupancy Memory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C2NOLGQT}},
  note         = {Machine review of arXiv:2607.05543}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Semantic occupancy provides a structured spatial memory for embodied indoor agents by jointly representing occupied regions, observed free space, unknown areas, and object semantics. However, existing indoor occupancy benchmarks and methods mainly focus on single-view prediction or room-level online perception, leaving long-horizon semantic mapping across connected indoor spaces underexplored. We introduce HIOcc, a hierarchical indoor occupancy benchmark that unifies ScanNet, ScanNet++, and Matterport3D under a common sparse semantic occupancy format while preserving their native observation geometries, including perspective RGB-D frames and pano-centric observation groups. HIOcc supports three complementary evaluation regimes: local semantic occupancy prediction, room-level online occupancy mapping, and building-level mapping across connected panoramic environments. We further propose GEM-Occ, a Gaussian Evidence Memory framework for semantic occupancy mapping. Rather than using pointmaps as persistent map states, GEM-Occ treats local visual geometry predictions as transient evidence, converts them into semantic Gaussian occupancy evidence and free-space ray evidence, and fuses them into a persistent hierarchical memory through visibility- and uncertainty-aware causal updates. The memory is organized into local caches, room-level submaps, and a building-level graph, and can be queried at any time through Gaussian-to-occupancy splatting. Experiments on HIOcc show that GEM-Occ improves local occupancy prediction, online map stability, free-space reasoning, revisit consistency, and building-level scalability over prior indoor occupancy and Gaussian-based mapping baselines.

Figures

Figures reproduced from arXiv: 2607.05543 by Baorui Peng, Bohan Li, Chang Wen Chen, Hongsi Liu, Hu Zhu, Mingqi Yuan, Wenjun Zeng, Xianda Guo, Xin Jin.

Figure 1
Figure 1. Figure 1: Overview of HIOcc and GEM-Occ. HIOcc provides hierarchical indoor semantic occupancy annotations across local views, rooms, and connected buildings. GEM-Occ builds persistent semantic Gaussian memory for local prediction, room-level online mapping, and building￾level mapping. Existing benchmarks reflect the same limitation. Foundational indoor datasets such as NYUv2, Scan￾Net, ScanNet++, and Matterport3D h… view at source ↗
Figure 2
Figure 2. Figure 2: HIOcc annotation pipeline. Calibrated observations and annotated meshes are converted into scene-level semantic voxels, cropped around each frame or panorama, visibility-filtered, and saved as sparse semantic occupancy annotations. spective and panoramic observations, and connected building-scale environments under a unified hierarchical online mapping formulation. In contrast, GEM-Occ studies hierarchical… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of GEM-Occ. The framework separates local evidence learning from streaming [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative local occupancy prediction. GEM-Occ produces cleaner semantic structure and fewer spurious occupied regions on HIOcc compared to other methods. but accumulated ray evidence supports free space, and unknown otherwise. We train the local encoder and Gaussian evidence adapter with view-conditioned supervision from HIOcc using L = λoccLocc + λsemLsem + λgeoLgeo + λrayLray. At test time, the update … view at source ↗
Figure 5
Figure 5. Figure 5: Streaming building-level (multi-room) mapping. GEM-Occ incrementally builds seman￾tic occupancy over connected panoramic observations. Ablation Studies [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional Building-level Occupancy Mapping [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. VGOcc: Learning Visual-Geometric Gaussians for Vision-Centric 3D Driving Occupancy Prediction

    cs.CV 2026-07 conditional novelty 5.0

    VGOcc fuses frozen VGGT/DINOv2 features with sparse 3D Gaussians to reach 34.07 SC IoU and 21.75 SSC mIoU on nuScenes, besting prior vision-only occupancy methods.

Reference graph

Works this paper leans on

54 extracted references · 6 linked inside Pith · cited by 1 Pith paper

  1. [1]

    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

  2. [2]

    Cao and R

    A.-Q. Cao and R. De Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, 2022

    2022

  3. [3]

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

    2023

  4. [4]

    Zhang, Z

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

    2023

  5. [5]

    H. Yu, Y . Wang, Y . Chen, and Z. Zhang. Monocular occupancy prediction for scalable indoor scenes. InECCV, 2024. 16

    2024

  6. [6]

    B. Li, Y . Sun, J. Dong, Z. Zhu, J. Liu, X. Jin, and W. Zeng. One at a time: Progressive multi-step volumetric probability learning for reliable 3d scene perception. InAAAI, 2024

    2024

  7. [7]

    J. Li, M. Lu, H. Wang, C. Gu, W. Zheng, L. Du, and S. Zhang. Sliceocc: Indoor 3d semantic occupancy prediction with vertical slice representation.ICRA, 2025

    2025

  8. [8]

    Huang, W

    Y .-K. Huang, W. Zheng, Y . Zhang, J. Zhou, and J. Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction.CVPR, 2023

    2023

  9. [9]

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

    2023

  10. [10]

    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. In ICCV, 2023

    2023

  11. [11]

    X. Tian, T. Jiang, L. Yun, Y . Mao, H. Yang, Y . Wang, Y . Wang, and H. Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. InNeurIPS Datasets and Benchmarks Track, 2023

    2023

  12. [12]

    Y . Li, S. Li, X. Liu, M. Gong, K. Li, N. Chen, Z. Wang, Z. Li, T. Jiang, F. Yu, Y . Wang, H. Zhao, Z. Yu, and C. Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. InIROS, 2024

    2024

  13. [13]

    C. Sima, W. Tong, T. Wang, L. Chen, S. Wu, H. Deng, Y . Gu, L. Lu, P. Luo, D. Lin, and H. Li. Scene as occupancy.arXiv, 2306.02851, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    J. Ma, X. Chen, J. Huang, J. Xu, Z. Luo, J. Xu, W. Gu, R. Ai, and H. Wang. Cam4DOcc: Benchmark for Camera-Only 4D Occupancy Forecasting in Autonomous Driving Applications. InCVPR, 2024

    2024

  15. [15]

    Contributors

    O. Contributors. Openscene: The largest up-to-date 3d occupancy prediction benchmark in autonomous driving.https://github.com/OpenDriveLab/OpenScene, 2023

    2023

  16. [16]

    W. Gan, N. Mo, H. Xu, and N. Yokoya. A comprehensive framework for 3d occupancy estimation in autonomous driving.IEEE Transactions on Intelligent Vehicles, 2024

    2024

  17. [17]

    Huang, W

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

    2024

  18. [18]

    Zhang, J

    C. Zhang, J. Yan, Y . Wei, J. Li, L. Liu, Y . Tang, Y . Duan, and J. Lu. Occnerf: Advancing 3d occupancy prediction in lidar-free environments.IEEE Transactions on Image Processing, 2025

    2025

  19. [19]

    D. Chen, H. Zheng, Y . Zhou, X. Li, W. Liao, T. He, P. Peng, and J. Shen. Semantic causality- aware vision-based 3d occupancy prediction.ICCV, 2025

    2025

  20. [20]

    B. Li, J. Deng, W. Zhang, Z. Liang, D. Du, X. Jin, and W. Zeng. Hierarchical temporal context learning for camera-based semantic scene completion. InECCV, 2024

    2024

  21. [21]

    B. Li, X. Jin, J. Wang, Y . Shi, Y . Sun, X. Wang, Z. Ma, B. Xie, C. Ma, X. Yang, et al. Occscene: Semantic occupancy-based cross-task mutual learning for 3d scene generation.TPAMI, 2025

    2025

  22. [22]

    B. Li, X. Jin, H. Zhu, H. Liu, R. Li, J. Guo, K. Cai, C. Ma, Y . Jin, H. Zhao, X. Yang, and W. Zeng. Scaling up occupancy-centric driving scene generation: Dataset and method.TPAMI, 2026

    2026

  23. [23]

    Cao and T.-H

    A.-Q. Cao and T.-H. Vu. Occany: Generalized unconstrained urban 3d occupancy. InCVPR, 2026. 17

    2026

  24. [24]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM TOG, 2023

    2023

  25. [25]

    Huang, W

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

    2024

  26. [26]

    Huang, A

    Y . Huang, A. Thammatadatrakoon, W. Zheng, Y . Zhang, D. Du, and J. Lu. Gaussianformer-2: Probabilistic gaussian superposition for efficient 3d occupancy prediction.CVPR, 2025

    2025

  27. [27]

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

    2025

  28. [28]

    L. Zhao, S. Wei, J. Hays, and L. Gan. Gaussianformer3d: Multi-modal gaussian-based semantic occupancy prediction with 3d deformable attention. InICRA, 2026

    2026

  29. [29]

    R. Qian, H. Cao, T. Deng, S. Yuan, and L. Xie. Splatssc: Decoupled depth-guided gaussian splatting for semantic scene completion. InAAAI, 2026

    2026

  30. [30]

    C. Zhou, Y . Luo, and C. Chen. Generalizing visual geometry priors to sparse gaussian occupancy prediction. InCVPR, 2026

    2026

  31. [31]

    Silberman, D

    N. Silberman, D. Hoiem, P. Kohli, and R. Fergus. Indoor segmentation and support inference from rgbd images. InECCV, 2012

    2012

  32. [32]

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner. Scannet: Richly- annotated 3d reconstructions of indoor scenes. InCVPR, 2017

    2017

  33. [33]

    Yeshwanth, Y .-C

    C. Yeshwanth, Y .-C. Liu, M. Nießner, and A. Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. InICCV, 2023

    2023

  34. [34]

    Chang, A

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y . Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In3DV, 2017

    2017

  35. [35]

    Y . Wu, W. Zheng, S. Zuo, Y . Huang, J. Zhou, and J. Lu. Embodiedocc: Embodied 3d occupancy prediction for vision-based online scene understanding. InICCV, 2025

    2025

  36. [36]

    S.-C. Wu, K. Tateno, N. Navab, and F. Tombari. Scfusion: Real-time incremental scene reconstruction with semantic completion. In3DV, 2020

    2020

  37. [37]

    T. Wang, X. Mao, C. Zhu, R. Xu, R. Lyu, P. Li, X. Chen, W. Zhang, K. Chen, T. Xue, X. Liu, C. Lu, D. Lin, and J. Pang. Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. InCVPR, 2024

    2024

  38. [38]

    B. Li, Y . Sun, Z. Liang, D. Du, Z. Zhang, X. Wang, Y . Wang, X. Jin, and W. Zeng. Bridging stereo geometry and bev representation with reliable mutual interaction for semantic scene completion.arXiv preprint arXiv:2303.13959, 2023

    Pith/arXiv arXiv 2023

  39. [39]

    H. Wang, X. Wei, X. Zhang, J. Li, C. Bai, Y . Li, M. Lu, W. Zheng, and S. Zhang. Embod- iedocc++: Boosting embodied 3d occupancy prediction with plane regularization and uncertainty sampler. InACM MM, 2025

    2025

  40. [40]

    Zhang, Q

    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

  41. [41]

    Y . Guo, S. Mentasti, X. Jin, M. Frosi, and M. Matteucci. Sgr-occ: Evolving monocular priors for embodied 3d occupancy prediction via soft-gating lifting and semantic-adaptive geometric refinement.arXiv preprint arXiv:2603.14076, 2026. 18

    arXiv 2026

  42. [42]

    W. Cui, H. Wang, W. Qin, Y . Guo, G. Han, W. Zhao, J. Cao, Z. Zhang, J. Zhong, J. Sun, P. Sun, S. Shi, B. Jiang, J. Ma, J. Wang, H. Cheng, Z. Liu, Y . Wang, Z. Zhu, G. Huang, J. Tang, and Q. Zhang. Humanoid occupancy: Enabling a generalized multimodal occupancy perception system on humanoid robots.arXiv preprint arXiv:2507.20217, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    H. Zhu, B. Li, X. Guo, Y . Peng, Z. Zhu, X. Jin, W. Zeng, and C. W. Chen. Bridging 3d gaussians and semantic occupancy for comprehensive open-vocabulary scene understanding from unposed images.preprint arXiv:2607.01633, 2026

    Pith/arXiv arXiv 2026

  44. [44]

    S. Wang, V . Leroy, Y . Cabon, B. Chidlovskii, and J. Revaud. Dust3r: Geometric 3d vision made easy. InCVPR, 2024

    2024

  45. [45]

    Leroy, Y

    V . Leroy, Y . Cabon, and J. Revaud. Grounding image matching in 3d with mast3r, 2024

    2024

  46. [46]

    Wang*, Y

    Q. Wang*, Y . Zhang*, A. Holynski, A. A. Efros, and A. Kanazawa. Continuous 3d perception model with persistent state. InCVPR, 2025

    2025

  47. [47]

    Y . Wu, W. Zheng, J. Zhou, and J. Lu. Point3r: Streaming 3d reconstruction with explicit spatial pointer memory. InNeurIPS, 2025

    2025

  48. [48]

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny. Vggt: Visual geometry grounded transformer. InCVPR, 2025

    2025

  49. [49]

    X. Chen, Y . Chen, Y . Xiu, A. Geiger, and A. Chen. Ttt3r: 3d reconstruction as test-time training. InICLR, 2026

    2026

  50. [50]

    Keetha, N

    N. Keetha, N. Müller, J. Schönberger, L. Porzi, Y . Zhang, T. Fischer, A. Knapitsch, D. Zauss, E. Weber, N. Antunes, J. Luiten, M. Lopez-Antequera, S. R. Bulò, C. Richardt, D. Ramanan, S. Scherer, and P. Kontschieder. Mapanything: Universal feed-forward metric 3d reconstruction. In3DV, 2026

    2026

  51. [51]

    Behley, M

    J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. Se- mantickitti: A dataset for semantic scene understanding of lidar sequences. InICCV, 2019

    2019

  52. [52]

    B. Li, S. Yang, B. Peng, X. Guo, E. Zhang, Y . Tao, J. Duan, D. Xu, Q. Dou, X. Jin, et al. From articulated kinematics to routed visual control for action-conditioned surgical video generation. arXiv preprint arXiv:2605.08712, 2026

    Pith/arXiv arXiv 2026

  53. [53]

    B. Li, Z. Ma, D. Du, B. Peng, Z. Liang, Z. Liu, C. Ma, Y . Jin, H. Zhao, W. Zeng, et al. Omninwm: Omniscient driving navigation world models. InECCV, 2026

    2026

  54. [54]

    B. Li, J. Deng, Y . Sun, X. Wang, X. Jin, and W. Zeng. Hierarchical context alignment with disentangled geometric and temporal modeling for semantic occupancy prediction.TPAMI, 2026. 19

    2026