Pith. sign in

REVIEW 4 major objections 3 minor 50 references

OccTrack360 establishes a benchmark and baseline for 4D panoptic occupancy tracking from surround-view fisheye cameras, with all-direction visibility masks.

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 ↗

A new benchmark and baseline method for 4D panoptic occupancy tracking with surround-view fisheye cameras, built from KITTI-360.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection OccTrack360 is a plausibly valuable fisheye benchmark, but the planar-XY backward completion and thin evaluation need fixes before I'd trust its numbers. the 4 major comments →

arxiv 2603.08521 v3 pith:WVD5IQJ2 submitted 2026-03-09 cs.CV cs.ROeess.IV

OccTrack360: 4D Panoptic Occupancy Tracking from Surround-View Fisheye Cameras

Yongzhi Lin , Kai Luo , Yuanfan Zheng , Hao Shi , Mengfei Duan , Yang Liu , Kailun Yang This is my paper
classification cs.CV cs.ROeess.IV
keywords 4D panoptic occupancy trackingfisheye cameraoccupancy benchmarkUnified Projection Modelocclusion maskfield-of-view maskspherical liftinginstance tracking
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.

The reading

The paper argues that existing occupancy benchmarks cannot measure long-horizon instance-level tracking under wide-field-of-view fisheye sensing, so it builds OccTrack360, a new benchmark with sequences of 174 to 2234 frames, instance-level voxel annotations, an occlusion mask covering all voxel directions, and a fisheye field-of-view mask derived from the Unified Projection Model. To give the benchmark a usable baseline, the paper proposes FoSOcc, a framework with a Center Focusing Module for instance-aware localization and a Fisheye-based Enhanced Lifting module that adapts 2D-to-3D feature lifting to fisheye distortion. On Occ3D-Waymo, FoSOcc improves segmentation quality for small and geometrically regular categories. On OccTrack360, FoSOcc raises occupancy segmentation scores but its tracking association metrics are lower than those of the prior TrackOcc baseline, a trade-off the paper reports explicitly.

Core claim

The paper claims that a valid 4D fisheye occupancy tracking benchmark needs both long temporal sequences and principled visibility supervision, and that fisheye-aware lifting plus center-focused supervision yields a strong baseline. The key supervision insight is that an occlusion mask must cast rays from every boundary voxel of the voxel space, not only from occupied voxels, so that upward and other sparse directions are not silently filtered out during training. The key geometric insight is that fisheye lifting should follow the Unified Projection Model, using the mirror parameter to convert image radius to a spherical angle, with a rectification bound that prevents invalid complex project

What carries the argument

The central objects are (1) the all-direction occlusion mask, which covers every direction in the voxel domain and avoids the incomplete ray coverage of prior masks, and (2) the MEI-based fisheye FoV mask, which computes the maximum radial angle from the Unified Projection Model and extends it across the voxel space to indicate valid fisheye supervision. On the method side, the Center Focusing Module replaces boundary-sensitive voxel offsets with a product-based focus feature that peaks at instance centers, and the Fisheye-based Enhanced Lifting module unprojects image features along fisheye rays using the mirror parameter and distortion coefficients.

Load-bearing premise

The ground-truth labels for occluded rear voxels are created by aligning earlier forward-facing voxels under planar XY motion and normalizing with SVD, which silently discards vertical motion; if real vertical variation is significant, the labels carry systematic vertical error.

What would settle it

Run the benchmark's label-generation pipeline on sequences with known uphill or downhill slopes and compare the completed rear voxels against dense LiDAR or manual annotation; if vertical error grows with the slope grade, the Z-truncating alignment is falsified.

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

If this is right

  • If OccTrack360 is adopted, researchers gain a common ground for comparing fisheye 4D occupancy trackers on long sequences, not just short clips.
  • The all-direction occlusion mask can prevent training-time filtering that hides voxel ambiguity, making predictions visually and geometrically interpretable.
  • The Fisheye-based Enhanced Lifting module provides a reusable recipe for adapting perspective lifting pipelines to wide-field-of-view cameras.
  • The reported tracking-metric trade-off suggests that instance association, not geometry alone, is the current bottleneck for fisheye 4D tracking.

Where Pith is reading between the lines

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

  • The backward-completion step could be transferred to pinhole benchmarks to add rear-object ground truth, but its planar motion assumption would need explicit handling for sloped roads.
  • The vertical error introduced by SVD normalization could be measured by comparing completed voxels against dense LiDAR on uphill and downhill segments; if it grows with grade, the labels carry systematic bias.
  • The FoV mask's manual truncation of the projected circle could be automated with a closed-form boundary polygon, making similar masks easier to build for other fisheye datasets.
  • The center-focus supervision idea could generalize to any dense 3D prediction task where boundary supervision is fragile under distortion, such as fisheye BEV segmentation.
Share X Bluesky LinkedIn Reddit HN

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 / 3 minor

Summary. The paper introduces OccTrack360, a benchmark for 4D panoptic occupancy tracking from surround-view fisheye cameras, built on KITTI-360, with long temporal sequences (174–2234 frames), instance-level voxel IDs, an all-direction occlusion mask, and an MEI-based fisheye field-of-view mask. The authors also propose FoSOcc, a fisheye-oriented occupancy tracking framework with a Center Focusing Module (CFM) for instance-center supervision and a Fisheye-based Enhanced Lifting (FEL) module that extends LSS lifting to the unified projection model. Experiments are reported on Occ3D-Waymo and OccTrack360, showing CFM gains on regular categories and establishing a baseline on the new benchmark.

Significance. If the benchmark construction is sound, OccTrack360 fills a real gap: no existing benchmark provides surround-view fisheye imagery, long sequences, and instance-level voxel tracking simultaneously. The proposed all-direction occlusion mask and MEI-based FoV mask are reasonable and more principled than prior pinhole-only visibility masks. The code and benchmark are promised to be public, which strengthens reproducibility. However, the paper's central claims rest on several load-bearing issues—a mathematical inconsistency in the FEL rectification bound, evaluation on a single sequence, a missing FEL ablation, and a known but unquantified vertical bias in the backward completion step—so the current evidence does not yet fully support the stated conclusions.

major comments (4)
  1. The rectification bound in Eq. (10) is inconsistent with the derivation in Eqs. (3)–(5). Eq. (5) correctly derives a_max = sqrt(1/(ξ²−1)) from the discriminant condition ∆ ≥ 0, but Eq. (10) uses arect = min(a, sqrt((1+ξ)/(ξ−1))). For ξ > 1, sqrt((1+ξ)/(ξ−1)) is always larger than sqrt(1/(ξ²−1)) by a factor of (ξ+1); e.g., for ξ=1.5 the former is ≈2.24 while the correct bound is ≈0.89. Thus the min() in Eq. (10) will not clamp a to the admissible range, and the subsequent inversion via Eq. (4) can produce complex cosθ values. This undermines the geometric validity of FEL, which is a core contribution. The manuscript must correct Eq. (10), specify which root of Eq. (4) is used, and retest the method with the corrected bound.
  2. The OccTrack360 evaluation is performed only on the Seq06 sequence, despite the benchmark claiming diverse sequences with 174–2234 frames. A single sequence cannot support claims about long-term or diverse evaluation, and it makes the reported numbers highly sensitive to one environmental layout, one set of traffic participants, and one slope profile. The authors should either evaluate on a larger subset of their benchmark (and report per-sequence variance) or substantially temper the diversity claim. This is load-bearing because the benchmark's central selling point is its long and varied sequences.
  3. FEL, one of the two primary contributions, is never ablated. Table III ablates only the CFM components (instance-level normalization and supervised focus feature) on Occ3D-Waymo, which uses pinhole cameras and does not involve FEL. No experiment isolates the contribution of FEL on OccTrack360. Without such an ablation, the improvements in Table II cannot be attributed to the fisheye lifting module. Additionally, Table II shows that FoSOcc decreases OccAQ (from 17.07 to 13.11 all-FoV; 22.32 to 17.02 Fisheyes) and OccSTQ (from 14.84 to 13.32 all-FoV; 17.20 to 15.71 Fisheyes), which contradicts the abstract's claim that the method 'improves occupancy tracking quality.' The authors should clarify whether the claim is about segmentation quality only, and add the missing FEL ablation.
  4. The backward completion of occluded voxels assumes planar XY motion and uses SVD normalization that, as the text admits, 'inevitably disregards genuine variations along the z-axis.' On sloped terrain (Fig. 4), this introduces a systematic vertical bias in the completed occupancy. Because the all-direction occlusion mask is constructed from this occupancy by ray-casting from boundary voxels, the mask inherits the same vertical error, and every benchmark metric (OccSQ/OccAQ/OccSTQ) is evaluated against a ground truth that can be vertically misaligned by more than the 0.2 m voxel size on inclines. The paper acknowledges the issue but does not quantify its magnitude or downstream impact. This is load-bearing for the benchmark's validity; the authors must either provide a quantitative analysis of the vertical error (e.g., on segments with known slopes) or adjust the completion procedure to ac
minor comments (3)
  1. The symbol R is introduced as 'the filling range' but then used as if it were a binary mask via the Hadamard product R◦V. This is confusing; please define R clearly (e.g., a 3D binary mask over the fill region).
  2. The lifting module is called 'Spherical Lift Module (SLM)' in the Introduction but 'Fisheye-based Enhanced Lifting (FEL)' in the abstract and Section IV. Please use one name consistently.
  3. The choice between the two roots for cosθ is not specified. The text says 'the chosen root' but never defines which root is used. Since this is central to FEL, please state the explicit closed form, including the sign before the square root.

Circularity Check

0 steps flagged

No significant circularity: benchmark and method are derived from external data and standard camera models; the admitted z-axis bias is a validity risk, not a circular step.

full rationale

I walked the claimed derivation chain. The OccTrack360 benchmark labels are constructed from KITTI-360 bounding-box and semantic annotations by deterministic voxelization, forward projection, backward completion, ray-cast occlusion masking, and an MEI-based fisheye FoV mask. These are data-construction procedures, not predictions derived from the method being evaluated. FoSOcc is trained with standard supervised losses on the resulting labels and evaluated both on the external Occ3D-Waymo benchmark and on OccTrack360; no fitted parameter is renamed as a prediction, and no equation is defined in terms of the metric it is claimed to explain. The paper's own admission that the planar-XY SVD normalization in backward completion 'inevitably disregards genuine variations along the z-axis' (Section III, Object Completion; Fig. 4) is a legitimate ground-truth validity concern, but it is a bias shared by the supervision and the evaluation, not a circular reduction: it does not force any method's score. Self-citations to prior work such as the voxel-offset formulation or TrackOcc's metrics are used as baselines or building blocks, not as load-bearing proofs. The central contributions—all-direction occlusion masks, MEI-based FoV masks, and fisheye-aware lifting—stand independently of these citations. Therefore the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The paper introduces no fitted free parameters or new physical entities. Its central claims rest on domain assumptions about the fidelity of KITTI-360 bounding boxes, the MEI camera model, and the planar-motion backward completion.

axioms (4)
  • domain assumption Backward completion of occluded voxels can be approximated by aligning current and previous forward-facing voxels under planar motion (rotation about Z, translation in XY), suppressing z-axis variation.
    Used in Section III Object Completion, Eq. (1)-(2); the paper acknowledges z errors and uses SVD to suppress them, which ignores genuine vertical variations.
  • domain assumption The fisheye cameras are exactly described by the MEI (Unified Projection Model) with a single mirror parameter ξ; if non-central or different distortion models apply, the FoV mask is invalid.
    Used in FoV mask construction Section III and FEL Section IV.C; MEI is an approximation.
  • domain assumption KITTI-360 3D bounding box tracks provide accurate instance-level dynamic poses for transforming voxels across frames.
    Relied on in Eq. (2) for dynamic voxel transformation; errors in box tracks propagate to labels.
  • domain assumption The occlusion mask computed from voxel boundaries is sufficient for visibility supervision; the boundary-voxel ray construction covers all directions without missing valid rays.
    Section III Visibility Constraint; this is a modeling choice.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of OccTrack360: 4D Panoptic Occupancy Tracking from Surround-View Fisheye Cameras." pith.science (2026). https://pith.science/paper/WVD5IQJ2

@misc{pith2026260308521,
  author       = {Pith},
  title        = {Pith review of: OccTrack360: 4D Panoptic Occupancy Tracking from Surround-View Fisheye Cameras},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WVD5IQJ2}},
  note         = {Machine review of arXiv:2603.08521}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Understanding dynamic 3D environments in a spatially continuous and temporally consistent manner is fundamental for robotics and autonomous driving. While recent advances in occupancy prediction provide a unified representation of scene geometry and semantics, progress in 4D panoptic occupancy tracking remains limited by the lack of benchmarks that support surround-view fisheye sensing, long temporal sequences, and instance-level voxel tracking. To address this gap, we present OccTrack360, a new benchmark for 4D panoptic occupancy tracking from surround-view fisheye cameras. OccTrack360 provides substantially longer and more diverse sequences (174~2234 frames) than prior benchmarks, together with principled voxel visibility annotations, including an all-direction occlusion mask and an MEI-based fisheye field-of-view mask. To establish a strong fisheye-oriented baseline, we further propose Focus on Sphere Occ (FoSOcc), a framework that addresses two core challenges in fisheye occupancy tracking: distorted spherical projection and inaccurate voxel-space localization. FoSOcc includes a Center Focusing Module (CFM) to enhance instance-aware spatial localization through supervised focus guidance, and a Fisheye-based Enhanced Lifting (FEL) that extends perspective lifting to fisheye imaging under the Unified Projection Model. Extensive experiments on Occ3D-Waymo and OccTrack360 show that our method improves occupancy tracking quality with notable gains on geometrically regular categories, and establishes a strong baseline for future research on surround-view fisheye 4D occupancy tracking. The benchmark and source code will be made publicly available at https://github.com/YouthZest-Lin/OccTrack360.

Figures

Figures reproduced from arXiv: 2603.08521 by Hao Shi, Kailun Yang, Kai Luo, Mengfei Duan, Yang Liu, Yongzhi Lin, Yuanfan Zheng.

Figure 1
Figure 1. Figure 1: Illustration of 4D panoptic occupancy tracking. Objects highlighted within the blue rectangles are the same objects. Our FoSOcc leverages fisheye images as input to perform comprehensive 4D panoptic occupancy tracking. unified interface for downstream tasks such as planning and interaction [3], [4]. Despite these advances, progress in 4D panoptic occu￾pancy tracking is still constrained by the lack of a su… view at source ↗
Figure 2
Figure 2. Figure 2: OccTrack360 facilitates comprehensive 4D panoptic occupancy prediction in 18 semantic classes using both pinhole [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Benchmark pipeline. Our benchmark integrates multiple inputs to derive the FoV mask, occlusion mask, and 4D centered occupancy labels. These masks are subsequently combined to filter the visible regions, yielding the final representation used for training and inference. “Gen” denotes generation. scene understanding, they typically overlook the distinct instance boundaries of static environmental elements—s… view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of z + error. The two frames shown above are both from sequence Seq00. The left and right subplots depict the 75th and 100th frames of SSCBench￾KITTI360 [6], respectively, corresponding to the 167th and 192th frames of the KITTI360 dataset [42]. The value z w represents the height translation extracted from the extrinsic parameters provided by KITTI360 [42]. At 75th frame, the ground-truth vox… view at source ↗
Figure 5
Figure 5. Figure 5: Limitation of the occlusion mask in Occ3D￾Waymo [5]. VM denotes the occlusion mask. Predictions without the occlusion mask exhibit voxel ambiguity, as certain directions were filtered out during training and thus excluded from back-propagation. represents the filling range. The transformation T ′ considers only rotation around the Z-axis and translation along the X- and Y-axes, since we aim to remain consi… view at source ↗
Figure 7
Figure 7. Figure 7: a ′ max in distorted image-plane. The red contour delineates the projection of the sampled radial parameter a onto the distorted image plane, following the application of the distortion and intrinsic mapping functions. calibrated using the Unified Projection Model (MEI) [7]. The mask is constructed in three steps. First, we determine the feasible projection limit of the undistorted normalized image-plane r… view at source ↗
Figure 8
Figure 8. Figure 8: Overall pipeline of FoSOcc. Our framework advances the standard occupancy pipeline through two novel contributions: (1) Fisheye-based Enhanced Lifting (FEL), which formulates the 2D-to-3D transformation within a spherical projection space by explicitly incorporating geometric priors—specifically the mirror parameter ξ, camera intrinsics, and distortion coefficients—to achieve high-fidelity spatial mapping;… view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of the Supervised Focus Feature (SFF). The heatmap provides a magnified view of the predicted semantic region. In the heatmap, red voxels represent higher digital values compared to the blue regions. TABLE I: Results on Occ3D-Waymo. G.O. denotes general objects. Methods are evaluated on categories with and without instance IDs. Best results are bolded, and ours is highlighted in green. OccSQ … 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

50 extracted references · 3 linked inside Pith

  1. [1]

    A survey on occupancy perception for autonomous driving: The information fusion perspective,

    H. Xuet al., “A survey on occupancy perception for autonomous driving: The information fusion perspective,”Information Fusion, 2025

    2025

  2. [2]

    A survey on end-to-end perception and prediction for autonomous driving,

    Y . Huet al., “A survey on end-to-end perception and prediction for autonomous driving,”Machine Intelligence Research, 2025

    2025

  3. [3]

    ALOcc: Adaptive lifting-based 3D semantic occu- pancy and cost volume-based flow predictions,

    D. Chenet al., “ALOcc: Adaptive lifting-based 3D semantic occu- pancy and cost volume-based flow predictions,” inICCV, 2025

    2025

  4. [4]

    V oxDet: Rethinking 3D semantic occupancy prediction as dense object detection,

    W. Liet al., “V oxDet: Rethinking 3D semantic occupancy prediction as dense object detection,” inNeurIPS, 2025

    2025

  5. [5]

    Occ3D: A large-scale 3D occupancy prediction benchmark for autonomous driving,

    X. Tianet al., “Occ3D: A large-scale 3D occupancy prediction benchmark for autonomous driving,” inNeurIPS, 2023

    2023

  6. [6]

    SSCBench: A large-scale 3D semantic scene completion benchmark for autonomous driving,

    Y . Liet al., “SSCBench: A large-scale 3D semantic scene completion benchmark for autonomous driving,” inIROS, 2024

    2024

  7. [7]

    Single view point omnidirectional camera calibration from planar grids,

    C. Mei and P. Rives, “Single view point omnidirectional camera calibration from planar grids,” inICRA, 2007

    2007

  8. [8]

    SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences,

    J. Behleyet al., “SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences,” inICCV, 2019

    2019

  9. [9]

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

    X. Wanget al., “OpenOccupancy: A large scale benchmark for surrounding semantic occupancy perception,” inICCV, 2023

    2023

  10. [10]

    Scene as occupancy,

    W. Tonget al., “Scene as occupancy,” inICCV, 2023

    2023

  11. [11]

    WoodScape: A multi-task, multi-camera fisheye dataset for autonomous driving,

    S. Yogamaniet al., “WoodScape: A multi-task, multi-camera fisheye dataset for autonomous driving,” inICCV, 2019

    2019

  12. [12]

    SynWoodScape: Synthetic surround-view fisheye camera dataset for autonomous driving,

    A. R. Sekkatet al., “SynWoodScape: Synthetic surround-view fisheye camera dataset for autonomous driving,”IEEE Robotics and Automa- tion Letters, 2022

    2022

  13. [13]

    TrackOcc: Camera-based 4D panoptic occupancy tracking,

    Z. Chenet al., “TrackOcc: Camera-based 4D panoptic occupancy tracking,” inICRA, 2025

    2025

  14. [14]

    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

  15. [15]

    Are we ready for autonomous driving? The KITTI vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? The KITTI vision benchmark suite,” inCVPR, 2012

    2012

  16. [16]

    Symphonize 3D semantic scene completion with contextual instance queries,

    H. Jianget al., “Symphonize 3D semantic scene completion with contextual instance queries,” inCVPR, 2024

    2024

  17. [17]

    GaussianOcc: Fully self-supervised and efficient 3D occupancy estimation with gaussian splatting,

    W. Ganet al., “GaussianOcc: Fully self-supervised and efficient 3D occupancy estimation with gaussian splatting,” inICCV, 2025

    2025

  18. [18]

    SA-Occ: Satellite-assisted 3D occupancy prediction in real world,

    C. Chenet al., “SA-Occ: Satellite-assisted 3D occupancy prediction in real world,” inICCV, 2025

    2025

  19. [19]

    MinkOcc: Towards real-time label-efficient semantic occupancy prediction,

    S. Szeet al., “MinkOcc: Towards real-time label-efficient semantic occupancy prediction,” inIROS, 2025

    2025

  20. [20]

    FB-OCC: 3D occupancy prediction based on forward- backward view transformation,

    Z. Liet al., “FB-OCC: 3D occupancy prediction based on forward- backward view transformation,”arXiv preprint arXiv:2307.01492, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Occupancy learning with spatiotemporal memory,

    Z. Lenget al., “Occupancy learning with spatiotemporal memory,” in ICCV, 2025

    2025

  22. [22]

    UniOcc: A unified benchmark for occupancy fore- casting and prediction in autonomous driving,

    Y . Wanget al., “UniOcc: A unified benchmark for occupancy fore- casting and prediction in autonomous driving,” inICCV, 2025

    2025

  23. [23]

    Cam4DOcc: Benchmark for camera-only 4D occupancy forecasting in autonomous driving applications,

    J. Maet al., “Cam4DOcc: Benchmark for camera-only 4D occupancy forecasting in autonomous driving applications,” inCVPR, 2024

    2024

  24. [24]

    Stream and query-guided feature aggregation for efficient and effective 3D occupancy prediction,

    S. Moonet al., “Stream and query-guided feature aggregation for efficient and effective 3D occupancy prediction,”arXiv preprint arXiv:2503.22087, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Latent gaussian splatting for 4D panoptic occupancy tracking,

    M. Luzet al., “Latent gaussian splatting for 4D panoptic occupancy tracking,”arXiv preprint arXiv:2602.23172, 2026

    Pith/arXiv arXiv 2026

  26. [26]

    PaSCo: Urban 3D panoptic scene completion with uncertainty awareness,

    A.-Q. Caoet al., “PaSCo: Urban 3D panoptic scene completion with uncertainty awareness,” inCVPR, 2024

    2024

  27. [27]

    PanoSSC: Exploring monocular panoptic 3D scene reconstruction for autonomous driving,

    Y . Shiet al., “PanoSSC: Exploring monocular panoptic 3D scene reconstruction for autonomous driving,” in3DV, 2024

    2024

  28. [28]

    Survey on fish-eye cameras and their applications in intelligent vehicles,

    Y . Qianet al., “Survey on fish-eye cameras and their applications in intelligent vehicles,”IEEE Transactions on Intelligent Transportation Systems, 2022

    2022

  29. [29]

    FisheyeBEVSeg: Surround view fisheye cam- eras based bird’s-eye view segmentation for autonomous driving,

    S. Yogamaniet al., “FisheyeBEVSeg: Surround view fisheye cam- eras based bird’s-eye view segmentation for autonomous driving,” in CVPRW, 2024

    2024

  30. [30]

    ArticuBEVSeg: Road semantic understanding and its application in bird’s eye view from panoramic vision system of long combination vehicles,

    W. Liu and W. Wang, “ArticuBEVSeg: Road semantic understanding and its application in bird’s eye view from panoramic vision system of long combination vehicles,”IEEE Robotics and Automation Letters, 2025

    2025

  31. [31]

    Dur360BEV: A real-world 360-degree single camera dataset and benchmark for bird-eye view mapping in autonomous driving,

    W. Eet al., “Dur360BEV: A real-world 360-degree single camera dataset and benchmark for bird-eye view mapping in autonomous driving,” inICRA, 2025

    2025

  32. [32]

    F2BEV: Bird’s eye view generation from surround-view fisheye camera images for automated driving,

    E. U. Samaniet al., “F2BEV: Bird’s eye view generation from surround-view fisheye camera images for automated driving,” inIROS, 2023

    2023

  33. [33]

    Exploring surround-view fisheye camera 3D object detection,

    C. Liet al., “Exploring surround-view fisheye camera 3D object detection,” inAAAI, 2026

    2026

  34. [34]

    OmniDet: Surround view cameras based multi-task visual perception network for autonomous driving,

    V . R. Kumaret al., “OmniDet: Surround view cameras based multi-task visual perception network for autonomous driving,”IEEE Robotics and Automation Letters, 2021

    2021

  35. [35]

    Surround-view fisheye BEV-perception for valet park- ing: Dataset, baseline and distortion-insensitive multi-task framework,

    Z. Wuet al., “Surround-view fisheye BEV-perception for valet park- ing: Dataset, baseline and distortion-insensitive multi-task framework,” IEEE Transactions on Intelligent V ehicles, 2023

    2023

  36. [36]

    Surround-view fisheye camera viewpoint augmentation for image semantic segmentation,

    J. Choet al., “Surround-view fisheye camera viewpoint augmentation for image semantic segmentation,”IEEE Access, 2023

    2023

  37. [37]

    DSBEV: Docking space segmentation for autonomous surface vehicle in bird’s eye view,

    C. Panget al., “DSBEV: Docking space segmentation for autonomous surface vehicle in bird’s eye view,”IEEE Transactions on Industrial Informatics, 2025

    2025

  38. [38]

    OmniOcc: Cylindrical voxel-based semantic occupancy prediction for omnidirectional vision systems,

    C. Wuet al., “OmniOcc: Cylindrical voxel-based semantic occupancy prediction for omnidirectional vision systems,”IEEE Access, 2025

    2025

  39. [39]

    EquivFisheye: A spherical fusion framework for panoramic 3D perception with surround-view fisheye cameras,

    Z. Yanget al., “EquivFisheye: A spherical fusion framework for panoramic 3D perception with surround-view fisheye cameras,”In- formation Fusion, 2025

    2025

  40. [40]

    FishBEV: Distortion-resilient bird’s eye view seg- mentation with surround-view fisheye cameras,

    H. Liet al., “FishBEV: Distortion-resilient bird’s eye view seg- mentation with surround-view fisheye cameras,”arXiv preprint arXiv:2509.13681, 2025

    arXiv 2025

  41. [41]

    KD360-V oxelBEV: LiDAR and 360-degree camera cross modality knowledge distillation for bird’s-eye-view segmenta- tion,

    Y . Sunet al., “KD360-V oxelBEV: LiDAR and 360-degree camera cross modality knowledge distillation for bird’s-eye-view segmenta- tion,”arXiv preprint arXiv:2512.15311, 2025

    arXiv 2025

  42. [42]

    KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D,

    Y . Liaoet al., “KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  43. [43]

    COTR: Compact occupancy transformer for vision- based 3D occupancy prediction,

    Q. Maet al., “COTR: Compact occupancy transformer for vision- based 3D occupancy prediction,” inCVPR, 2024

    2024

  44. [44]

    Close-range camera calibration,

    C. B. Duane, “Close-range camera calibration,”Photogramm. Eng, 1971

    1971

  45. [45]

    MinVIS: A minimal video instance segmentation framework without video-based training,

    D.-A. Huang, Z. Yu, and A. Anandkumar, “MinVIS: A minimal video instance segmentation framework without video-based training,” in NeurIPS, 2022

    2022

  46. [46]

    CTVIS: Consistent training for online video instance segmentation,

    K. Yinget al., “CTVIS: Consistent training for online video instance segmentation,” inICCV, 2023

    2023

  47. [47]

    3D multi-object tracking: A baseline and new evaluation metrics,

    X. Wenget al., “3D multi-object tracking: A baseline and new evaluation metrics,” inIROS, 2020

    2020

  48. [48]

    4D panoptic LiDAR segmentation,

    M. Aygunet al., “4D panoptic LiDAR segmentation,” inCVPR, 2021

    2021

  49. [49]

    Deep residual learning for image recognition,

    K. Heet al., “Deep residual learning for image recognition,” inCVPR, 2016

    2016

  50. [50]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inICLR, 2019

    2019

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.