Pith. sign in

REVIEW 4 major objections 7 minor 42 references

GTAD: Global Temporal Aggregation Denoising Learning for 3D Semantic Occupancy Prediction

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read GTAD claims camera-only occupancy prediction reaches state-of-the-art by aggregating global temporal history with an in-model denoising network.

desk verdict GTAD is a solid empirical architecture paper for temporal fusion in occupancy prediction, but the SOTA claim is unsupported and the average-pooled global queue raises an untested vertical-cue concern. read the letter →

arxiv 2507.20963 v1 pith:2QJ72KIS submitted 2025-07-28 cs.CV

Tianhao Li , Yang Li , Mengtian Li , Yisheng Deng , Weifeng Ge This is my paper
classification cs.CV
keywords 3Dsemanticoccupancypredictiontemporalaggregationdenoisinglearningbird's-eyeviewautonomousdrivingcamera-basedperceptionglobalinteractionnuScenes
verification ladder T0 review T1 audit T2 compute T3 formal

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 camera-based 3D occupancy prediction improves when a model aggregates history at the sequence level, not just between adjacent frames. It introduces GTAD, a network that keeps a queue of past voxel features, compresses them to bird's-eye-view maps, lets those maps interact through forward-backward attention, and then uses an in-model denoising network to combine this global condition with locally fused current features. The reported result is 40.76 mIoU on Occ3D-nuScenes after 12 training epochs, above PanoOcc trained for 12 epochs (36.63) and approaching its 24-epoch version (42.13), with lower memory than the 24-epoch model. If true, it suggests a cheaper training schedule can reach competitive occupancy accuracy by putting temporal information to work globally.

What carries the argument

The load-bearing object is the in-model latent denoising network combined with the Global Interaction Attention Module (GIAM). The denoising network, built on DeTrack's design, adds Gaussian noise to the current BEV feature map and iteratively subtracts predicted noise across several denoising blocks, conditioned on locally aligned voxel features and globally aggregated BEV history; GIAM is a cross-attention fusion between adjacent BEV frames in a temporal queue, with learnable time-decay weights, run once backward and once forward so every historical frame absorbs context from both sides. This machinery is what turns a short training schedule into competitive occupancy accuracy.

What would settle it

Train GTAD with the global temporal encoder removed, keeping only local temporal aggregation and denoising, on Occ3D-nuScenes at 12 epochs; if mIoU stays near 40.76, the claimed global aggregation is not the driver. Alternatively, train both GTAD and PanoOcc for 24 epochs; if GTAD still trails PanoOcc's 42.13, the state-of-the-art claim depends on the shorter schedule rather than the mechanism.

Watch

Extended reading notes

Core claim

GTAD's central claim is that global temporal aggregation—using a whole historical queue rather than one or two past frames—is the missing ingredient in camera-based 3D occupancy prediction, and that an in-model latent denoising network is an effective and memory-efficient way to inject that global context. The paper shows that its Global Interaction Attention Module, which exchanges BEV features across the queue in backward and forward passes with time-decay weights, contributes most of the gain: replacing it with temporal self-attention or plain concatenation drops mIoU from 40.76 to 30.13 or 34.25 respectively. The denoising network then removes Gaussian noise added to the current BEV map over several denoising blocks, conditioned on local and global temporal features, in a single forward pass.

Load-bearing premise

The global temporal branch assumes that average-pooling the voxel history into BEV maps keeps everything the denoising network needs, so if height-specific cues are lost in that compression, the global condition cannot recover them.

Editorial extensions

If this is right

  • If correct, camera-only occupancy models can reach near-24-epoch accuracy with only 12 training epochs, cutting training cost while retaining accuracy.
  • Global temporal aggregation can be integrated into existing BEV-based perception pipelines without the memory cost of keeping full-voxel historical queues.
  • The time-decay forward-backward interaction scheme applies beyond occupancy to other long-horizon perception tasks that need sequence context.
  • The denoising conditioning scheme recasts occupancy prediction as a conditional denoising problem solved inside a single forward pass, avoiding the multi-pass cost of diffusion sampling.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because voxel-to-BEV compression averages over height, objects whose occupancy signal is elevation-specific, such as overhangs and poles, may be underrepresented in the global condition; a multi-height BEV or selective pooling could test this directly.
  • The reported margin over PanoOcc holds only at 12 epochs versus PanoOcc's 24-epoch checkpoint; a matched 24-epoch comparison would clarify whether global aggregation genuinely closes the gap or mainly accelerates convergence.
  • The denoising block is task-agnostic, so the same global-temporal conditioning could transfer to video-based panoptic or instance tracking, where re-identifying occluded objects benefits from long-range history.
  • The corruption-scale sweet spot at t=800 suggests a tradeoff between exploration and fidelity; scheduling t across training rather than fixing it might further improve results.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes GTAD, a camera-based 3D semantic occupancy prediction framework that combines a spatial encoder, a local temporal encoder, a global temporal encoder with a Global Interaction Attention Module (GIAM), and an in-model latent denoising network adapted from DeTrack. The global encoder compresses a voxel feature queue into BEV maps via average pooling and applies backward-forward cross-attention with time-decay weights. The denoising network takes the locally aggregated voxel features and the globally aggregated BEV features as conditions, and progressively removes noise from a corrupted BEV feature map in a single forward pass. Experiments on Occ3D-nuScenes report 40.76 mIoU at 12 training epochs, and on nuScenes LiDAR semantic segmentation report 70.2 mIoU with an R50 backbone. Ablations cover the local encoder, global encoder fusion, number of denoising steps, and corruption scale.

Significance. If corroborated, GTAD would be a competitive temporal-aggregation design for short training schedules and a novel application of in-model latent denoising to dense 3D occupancy prediction. The paper ships a complete architecture, internally consistent ablations (Tables III–VI), and a memory comparison (Table VII). The interaction between local and global temporal conditions is a useful design axis, and the GIAM fusion is clearly better than concatenation or temporal self-attention in the reported setting. However, the state-of-the-art claim is not supported by Table I because the 24-epoch PanoOcc baseline is higher than the reported GTAD result, and the contribution of the global temporal aggregation is not fully isolated from the choices of pooling and denoising depth. The empirical margins in Table II are small and reported without uncertainty estimates.

major comments (4)
  1. [Section IV.D, Table I] The paper claims state-of-the-art performance for camera-based occupancy field prediction, but Table I shows PanoOcc trained for 24 epochs at 42.13 mIoU, which is higher than GTAD's 40.76 mIoU at 12 epochs. The abstract and contribution bullets claim SOTA without qualification, while the text acknowledges the shortfall to 24-epoch PanoOcc. The claim as stated is therefore not supported by the reported numbers. The authors should either train GTAD for 24 epochs and compare on equal footing, or explicitly restrict the claim to 12-epoch training schedules and revise the abstract and Section I accordingly.
  2. [Section III.C, Eqs. (5)–(6), Table V] The global temporal queue compresses the aligned voxel feature volume into BEV maps by average pooling along the Z axis, and the GIAM then interacts only on these pooled maps. Average pooling is many-to-one: it discards vertical structure, so the global condition cannot distinguish, for example, a tall vehicle from a low obstacle at the same footprint. Since the global path is the only source of long-range historical context in the design, information lost here cannot be recovered downstream. The ablations in Table V vary only the fusion module (Cat, TSA, GIAM), not the pooling operator or the representation dimensionality. A height-retaining baseline (for example, per-height-slice interaction, max pooling, or retaining a small number of height channels) is needed to verify that the global temporal aggregation actually delivers the vertical information that occupancy labels require.
  3. [Section III.D, Eq. (7)–(11), Table IV] The ablation on denoising steps in Table IV increases the number of denoising blocks from 1 to 9 and shows a peak at 6 steps. However, there is no control with the same block stack but without noise corruption and without the noise-prediction objective. The performance improvement from adding blocks could therefore be due to increased model capacity or deeper feature interactions rather than to the denoising formulation itself. To support the central claim that in-model latent denoising is the mechanism behind the gain, the authors should compare against a non-denoising stack of the same depth and width (for example, the same cross-attention and FFN layers with no noise added and no epsilon prediction).
  4. [Section IV.D, Table II] The LiDAR semantic segmentation results that support the superiority claim are reported as single runs without standard deviations or multiple seeds. The margins are small (GTAD 70.2 vs. TPVFormer-Base 68.9 vs. PanoOcc-Small-T 68.1), and Table II mixes backbones (R101-DCN for some methods, R50 for others). Without variance information, these differences cannot be distinguished from training noise. Reporting at least three seeds with mean and std for the main comparisons would substantially strengthen the empirical claim.
minor comments (7)
  1. [Eq. (7)] Equation (7) is malformed: `Softmax(q_xi k_t / sqrt(d) v_t)` mixes the attention logits with the value vector inside the softmax. It should be `Softmax(q_xi k_t^T / sqrt(d)) v_t` or an explicit attention-weight notation.
  2. [Eqs. (10)–(11)] The subscripts in Eqs. (10) and (11) are corrupted: `xi− I/l = x''_i − ϵ` and `x0 = xI − sum_j ϵ_j` use an undefined `I` and a slash. These should read `x_{i-1} = x''_i − ε` and `x_0 = x_I − Σ ε_j` (or the intended notation should be defined).
  3. [Section I] The reference list for BEV semantic segmentation in the first paragraph is `[1], [1], [2]`; the duplicate `[1]` should be corrected, and the second citation likely refers to a different work.
  4. [Section III.C] The symbol C is used both for the channel dimension in `B_t ∈ R^{H×W×C}` and for the number of semantic classes in Section III.A. Use a different symbol, such as D or C_feat, for the channel dimension to avoid ambiguity.
  5. [Section III.C] There is a typo in the heading text: `communcication` should be `communication`.
  6. [Section III.D and Table VI] The manuscript does not state the corruption scale t used in the main model or how t is sampled during training. Table VI reports a best value of t = 800, but the training procedure (e.g., fixed t, random sampling, schedule) is not described. Please specify the noise schedule and the exact t used for the reported results.
  7. [Section IV.D] The text says the model is evaluated on both the nuScenes test and validation sets, but Table II only reports validation-set results. Either add the test results or remove the mention of the test set.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GTAD is an empirical architecture paper whose benchmark results are externally evaluated; the DeTrack self-citation supplies a component, not a forced derivation.

full rationale

The paper's central claims are architectural and empirical: it introduces a global temporal aggregation encoder and an in-model latent denoising network for 3D semantic occupancy prediction, then evaluates them on Occ3D-nuScenes and nuScenes. No parameter is fitted to a target metric and then reported as a prediction; the reported mIoU values come from training on standard splits. The closest candidate for circularity is the self-citation of DeTrack [21], which provides the in-model latent denoising block. However, that citation supplies a reusable architectural component, not an unverified uniqueness theorem, and the present paper's occupancy application, GIAM module, and benchmark results are not derived from DeTrack by construction. The GIAM equations (Eqs. 5-6) define an explicit backward-forward interaction mechanism, and the denoising blocks (Eqs. 7-11) are a standard attention-plus-noise-prediction structure; neither reduces the output to its input by definition. The average-pooling compression to BEV is a representation choice and a possible performance bottleneck, but it is not a circular step: the claimed results are not forced by that pooling. The only self-citation is therefore not load-bearing, so no significant circularity is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The architecture introduces no new physical entities. The central claim rests on validation-tuned hyperparameters and on modeling assumptions inherited from prior work such as deformable attention and BEV compression.

free parameters (4)
  • denoising_blocks_l = 6
    The ablation in Table IV shows mIoU peaks at 6 denoising blocks; this is selected based on validation performance.
  • corruption_scale_t = 800
    Table VI reports mIoU peaks at t=800 on validation, so this hyperparameter is tuned on the validation set.
  • num_historical_frames_k = 4
    Section IV.D states temporal fusion uses 4 frames, but no ablation for k is shown.
  • loss_weights_lambda_1_to_5 = not reported
    Equations 12-13 define lambda_1..lambda_5 but no values are given; they are either inherited from prior work or tuned implicitly.
assumptions (4)
  • domain assumption In-model latent denoising with l blocks approximates multi-step diffusion denoising in one forward pass.
    Adopted from DeTrack [21] without proof, used in Section III.D.
  • domain assumption BEV average pooling preserves sufficient 3D structure for global temporal aggregation.
    Section III.C states voxel representations are impractical for global queues and compresses them to BEV.
  • domain assumption Ego-pose alignment of historical voxel features is accurate.
    The local temporal encoder uses voxel alignment in 3D space, which requires reliable pose information from nuScenes.
  • standard math Training with LiDAR sparse semantic labels and occupancy labels yields correct voxel supervision.
    The supervision setup follows PanoOcc and Occ3D benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GTAD: Global Temporal Aggregation Denoising Learning for 3D Semantic Occupancy Prediction." pith.science (2026). https://pith.science/paper/2QJ72KIS

@misc{pith2026250720963,
  author       = {Pith},
  title        = {Pith review of: GTAD: Global Temporal Aggregation Denoising Learning for 3D Semantic Occupancy Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2QJ72KIS}},
  note         = {Machine review of arXiv:2507.20963}
}
read the original abstract

Accurately perceiving dynamic environments is a fundamental task for autonomous driving and robotic systems. Existing methods inadequately utilize temporal information, relying mainly on local temporal interactions between adjacent frames and failing to leverage global sequence information effectively. To address this limitation, we investigate how to effectively aggregate global temporal features from temporal sequences, aiming to achieve occupancy representations that efficiently utilize global temporal information from historical observations. For this purpose, we propose a global temporal aggregation denoising network named GTAD, introducing a global temporal information aggregation framework as a new paradigm for holistic 3D scene understanding. Our method employs an in-model latent denoising network to aggregate local temporal features from the current moment and global temporal features from historical sequences. This approach enables the effective perception of both fine-grained temporal information from adjacent frames and global temporal patterns from historical observations. As a result, it provides a more coherent and comprehensive understanding of the environment. Extensive experiments on the nuScenes and Occ3D-nuScenes benchmark and ablation studies demonstrate the superiority of our method.

Figures

Figures reproduced from arXiv: 2507.20963 by the authors.

Figure 1
Figure 1. Comparison of temporal aggregation methods. Our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. GTAD framework starts with an image backbone network to extract multi-scale features from multi-view images [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The Global Temporal Information Interaction Module [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Our method is compared with the ground truth of Occ3d-nuScenes through visualizations in various scenarios. In addition to urban areas, we selected challenging scenarios. The results demonstrate that our method performs well across different scenarios. Occ [22], are tr…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 15 canonical work pages

  1. [1]

    Cross-view transformers for real-time map-view semantic segmentation,

    B. Zhou and P. Kr ¨ahenb¨uhl, “Cross-view transformers for real-time map-view semantic segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 13 760–13 769

    2022

  2. [2]

    Lidar panoptic segmentation for autonomous driving,

    A. Milioto, J. Behley, C. McCool, and C. Stachniss, “Lidar panoptic segmentation for autonomous driving,” in 2020 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS) . IEEE, 2020, pp. 8505–8512

    work page 2020

  3. [3]

    Fiery: future instance prediction in bird’s- eye view from surround monocular cameras,

    A. Hu, Z. Murez, N. Mohan, S. Dudas, J. Hawke, V . Badrinarayanan, R. Cipolla, and A. Kendall, “Fiery: future instance prediction in bird’s- eye view from surround monocular cameras,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 15 273–15 282

    work page 2021

  4. [4]

    Is pseudo- lidar needed for monocular 3d object detection?

    D. Park, R. Ambrus, V . Guizilini, J. Li, and A. Gaidon, “Is pseudo- lidar needed for monocular 3d object detection?” in ICCV, 2021

    work page 2021

  5. [5]

    MMDetection3D: OpenMMLab next-generation platform for general 3D object detection,

    M. Contributors, “MMDetection3D: OpenMMLab next-generation platform for general 3D object detection,” https://github.com/ open-mmlab/mmdetection3d, 2020

    2020

  6. [6]

    Maptr: Structured modeling and learning for online vectorized hd map construction,

    B. Liao, S. Chen, X. Wang, T. Cheng, Q. Zhang, W. Liu, and C. Huang, “Maptr: Structured modeling and learning for online vectorized hd map construction,” arXiv preprint arXiv:2208.14437 , 2022

    arXiv 2022

  7. [7]

    Predicting semantic map representations from images using pyramid occupancy networks,

    T. Roddick and R. Cipolla, “Predicting semantic map representations from images using pyramid occupancy networks,” in CVPR, 2020

    work page 2020

  8. [8]

    Spatial Pruned Sparse Convolution for Efficient 3D Object Detection

    J. Liu, Y . Chen, X. Ye, Z. Tian, X. Tan, and X. Qi, “Spatial pruned sparse convolution for efficient 3d object detection,” arXiv preprint arXiv:2209.14201, 2022

    work page Pith review arXiv 2022

Show all 42 references
  1. [9]

    Multi- scale interaction for real-time lidar data segmentation on an embedded platform,

    S. Li, X. Chen, Y . Liu, D. Dai, C. Stachniss, and J. Gall, “Multi- scale interaction for real-time lidar data segmentation on an embedded platform,” IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 738–745, 2021

    2021

  2. [10]

    Occdepth: A depth-aware method for 3d semantic scene completion,

    R. Miao, W. Liu, M. Chen, Z. Gong, W. Xu, C. Hu, and S. Zhou, “Occdepth: A depth-aware method for 3d semantic scene completion,” arXiv preprint arXiv:2302.13540 , 2023

    2023 arXiv

  3. [11]

    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,” arXiv preprint arXiv:2302.07817, 2023

    2023 arXiv

  4. [12]

    Openoccupancy: A large scale benchmark for surround- ing semantic occupancy perception,

    X. Wang, Z. Zhu, W. Xu, Y . Zhang, Y . Wei, X. Chi, Y . Ye, D. Du, J. Lu, and X. Wang, “Openoccupancy: A large scale benchmark for surround- ing semantic occupancy perception,” arXiv preprint arXiv:2303.03991, 2023

    2023 arXiv

  5. [13]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 9087–9098

    2023

  6. [14]

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

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

    2022

  7. [15]

    Let occ flow: Self-supervised 3d occupancy flow prediction,

    Y . Liu, L. Mou, X. Yu, C. Han, S. Mao, R. Xiong, and Y . Wang, “Let occ flow: Self-supervised 3d occupancy flow prediction,” ArXiv, vol. abs/2407.07587, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:271088768

    2024 arXiv

  8. [16]

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

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

  9. [17]

    Henet: Hybrid encoding for end-to-end multi-task 3d perception from multi-view cameras,

    Z. Xia, Z. Lin, X. Wang, Y . Wang, Y . Xing, S. Qi, N. Dong, and M.-H. Yang, “Henet: Hybrid encoding for end-to-end multi-task 3d perception from multi-view cameras,” 2024. [Online]. Available: https://arxiv.org/abs/2404.02517

    2024 arXiv

  10. [18]

    Bevdet4d: Exploit temporal cues in multi-camera 3d object detection,

    J. Huang and G. Huang, “Bevdet4d: Exploit temporal cues in multi-camera 3d object detection,” 2022. [Online]. Available: https://arxiv.org/abs/2203.17054

    2022 arXiv

  11. [20]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” 2020. [Online]. Available: https://arxiv.org/abs/2006.11239

    2020 arXiv

  12. [21]

    Detrack: In-model latent denoising learning for visual object tracking,

    X. Zhou, J. Li, L. Hong, K. Jiang, P. Guo, W. Ge, and W. Zhang, “Detrack: In-model latent denoising learning for visual object tracking,” 2025. [Online]. Available: https://arxiv.org/abs/2501.02467

    2025 arXiv

  13. [22]

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

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

    2023 arXiv

  14. [23]

    Monoscene: Monocular 3d semantic scene completion,

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

    2022

  15. [24]

    Deformable detr: Deformable transformers for end-to-end object detection,

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

    2010 arXiv

  16. [25]

    Drinet++: Ef- ficient voxel-as-point point cloud segmentation,

    M. Ye, R. Wan, S. Xu, T. Cao, and Q. Chen, “Drinet++: Ef- ficient voxel-as-point point cloud segmentation,” arXiv preprint arXiv:2111.08318, 2021

    2021 arXiv

  17. [26]

    Search- ing efficient 3d architectures with sparse point-voxel convolution,

    H. Tang, Z. Liu, S. Zhao, Y . Lin, J. Lin, H. Wang, and S. Han, “Search- ing efficient 3d architectures with sparse point-voxel convolution,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII . Springer, 2020, pp. 685–702

    2020

  18. [27]

    Deepvoxels: Learning persistent 3d feature embeddings,

    V . Sitzmann, J. Thies, F. Heide, M. Nießner, G. Wetzstein, and M. Zoll- hofer, “Deepvoxels: Learning persistent 3d feature embeddings,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 2437–2446

    2019

  19. [28]

    Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo,

    Y . Li, H. Bao, Z. Ge, J. Yang, J. Sun, and Z. Li, “Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo,” arXiv preprint arXiv:2209.10248 , 2022

    2022 arXiv

  20. [29]

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

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

    2021 arXiv

  21. [30]

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

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

    2022 arXiv

  22. [31]

    Fb-bev: Bev representation from forward-backward view transforma- tions,

    Z. Li, Z. Yu, W. Wang, A. Anandkumar, T. Lu, and J. M. Alvarez, “Fb-bev: Bev representation from forward-backward view transforma- tions,” in ICCV, 2023, pp. 6919–6928

    2023

  23. [32]

    Bev-lanedet: a simple and effective 3d lane detection baseline,

    R. Wang, J. Qin, K. Li, Y . Li, D. Cao, and J. Xu, “Bev-lanedet: a simple and effective 3d lane detection baseline,” in CVPR, 2023

    2023

  24. [33]

    Bevpoolv2: A cutting-edge implementa- tion of bevdet toward deployment,

    J. Huang and G. Huang, “Bevpoolv2: A cutting-edge implementa- tion of bevdet toward deployment,” arXiv preprint arXiv:2211.17111 , 2022

    2022 arXiv

  25. [34]

    Fast-bev: Towards real-time on-vehicle bird’s-eye view perception,

    B. Huang, Y . Li, E. Xie, F. Liang, L. Wang, M. Shen, F. Liu, T. Wang, P. Luo, and J. Shao, “Fast-bev: Towards real-time on-vehicle bird’s-eye view perception,” arXiv preprint arXiv:2301.07870 , 2023

    2023 arXiv

  26. [35]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2980–2988

    2017

  27. [36]

    The lov ´asz-softmax loss: A tractable surrogate for the optimization of the intersection- over-union measure in neural networks,

    M. Berman, A. R. Triki, and M. B. Blaschko, “The lov ´asz-softmax loss: A tractable surrogate for the optimization of the intersection- over-union measure in neural networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 4413–4421

    2018

  28. [37]

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

    2020

  29. [38]

    Internimage: Exploring large-scale vision foundation models with deformable convolutions,

    W. Wang, J. Dai, Z. Chen, Z. Huang, Z. Li, X. Zhu, X. Hu, T. Lu, L. Lu, H. Li, et al. , “Internimage: Exploring large-scale vision foundation models with deformable convolutions,” arXiv preprint arXiv:2211.05778, 2022

    2022 arXiv

  30. [39]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  31. [40]

    Improving deep neural networks using softplus units,

    H. Zheng, Z. Yang, W. Liu, J. Liang, and Y . Li, “Improving deep neural networks using softplus units,” in 2015 International joint conference on neural networks (IJCNN) . IEEE, 2015, pp. 1–4

    2015

  32. [41]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101 , 2017

    2017 arXiv

  33. [42]

    Fcos3d: Fully convolutional one-stage monocular 3d object detection,

    T. Wang, X. Zhu, J. Pang, and D. Lin, “Fcos3d: Fully convolutional one-stage monocular 3d object detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 913–922

    2021

  34. [2023]

    Available: https://arxiv.org/abs/2306.10013

    [Online]. Available: https://arxiv.org/abs/2306.10013

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.