Pith. sign in

REVIEW 2 major objections 5 minor 43 references

OnlineBEV: Recurrent Temporal Fusion in Bird's Eye View Representations for Multi-Camera 3D Perception

T0 review · 2 major / 5 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 OnlineBEV argues that a single recurrent BEV feature, aligned by motion-guided deformable attention, can beat parallel multi-frame fusion and reaches 63.9% NDS on the nuScenes test set.

desk verdict The architecture is interesting, but the paper's own ablation contradicts its headline SOTA numbers, so the central empirical claim is unsupported. read the letter →

arxiv 2507.08644 v1 pith:PNUWYHYN submitted 2025-07-11 cs.CV

Junho Koh , Youngwoo Lee , Jungho Kim , Dongyoung Lee , Jun Won Choi This is my paper
classification cs.CV
keywords OnlineBEVrecurrenttemporalfusionbird'seyeviewmotion-guideddeformableattention3DobjectdetectionconsistencylossnuScenesmulti-cameraperception
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 sets out to close the gap between parallel temporal fusion, which stores a window of past bird's-eye-view (BEV) features, and recurrent fusion, which keeps one historical feature and updates it over time. It argues that recurrent fusion only fails because historical features drift out of alignment with current ones as objects move, and that this misalignment can be learned and corrected rather than avoided by keeping many frames. OnlineBEV's Motion-Guided BEV Fusion Network aligns the stored historical feature before fusing it, and a heatmap-based consistency loss explicitly supervises that alignment. On nuScenes, OnlineBEV reports 63.9% NDS and 55.8% mAP on the test set, outperforming SOLOFusion, StreamPETR, and SparseBEV, while using a single memory feature instead of a 17-frame bank. If the claim holds, long-history temporal fusion becomes nearly memory-free for dense BEV methods.

What carries the argument

The machine that carries the argument is the Motion-Guided BEV Fusion Network (MBFNet), composed of two parts. The Motion Feature Extractor (MFE) computes a motion context feature as channel-wise attention applied to the difference between historical and current BEV features (Eq. 1); static content cancels out while moving objects leave a signal. The Motion-Guided BEV Warping Attention (MGWA) then runs deformable attention in which the motion feature supplies the sampling offsets and attention weights, effectively moving the historical feature to where the current feature indicates objects now are. Around this sits the Heatmap-based Temporal Consistency Loss, which compares object-center heatmaps of the aligned history and the current frame with gradients blocked on the current side, so that training explicitly pulls the aligned history toward the present.

What would settle it

A decisive check would be to replace the learned motion features with ground-truth per-object velocity or optical-flow-derived offsets during alignment and compare on the nuScenes validation set; if the supervised motion signal does not beat the learned difference signal by a large margin, then the channel-difference guidance is not doing the alignment work that the paper claims. A second check is to evaluate on a subset of frames with fast-moving objects and see whether the reported gains over SOLOFusion persist when displacements are large.

Watch

Extended reading notes

Core claim

OnlineBEV's central claim is that a single recurrently maintained BEV feature can outperform a multi-frame memory bank, provided the stored feature is aligned with the current frame before fusion. Alignment is performed by MBFNet: a Motion Feature Extractor forms a motion map from the channel-wise difference between historical and current BEV features, and a Motion-Guided BEV Warping Attention module uses deformable attention to sample from the historical feature at offsets and weights produced by that motion map. The aligned historical feature is fused with the current one, the result becomes the new memory, and a heatmap-based temporal consistency loss pushes the aligned history to resemble the current frame during training. The paper reports that with a V2-99 backbone this pipeline obtains 55.8% mAP and 63.9% NDS on the nuScenes test split, surpassing the previous camera-only state of the art, and that the same framework improves BEV segmentation and 3D occupancy prediction.

Load-bearing premise

The whole gain rests on the assumption that the channel-wise difference between the historical BEV feature and the current one is a reliable guide to where and how much each feature should move when aligning history to the present.

Editorial extensions

If this is right

  • A recurrent design with one stored BEV feature can replace a 17-frame parallel memory bank at better accuracy, reducing memory use from 3.9 GB to 3.4 GB.
  • Dynamic-object misalignment, the main obstacle to long-range temporal fusion, can be handled by deformable attention guided by motion differences rather than by explicit optical flow.
  • The gain transfers beyond 3D detection: OnlineBEV raises BEV segmentation mIoU by 2.9 points and 3D occupancy mIoU by 1.2 points over SOLOFusion on nuScenes val.
  • Under synthetic motion blur and occlusion, OnlineBEV degrades less than SOLOFusion, widening its mAP and NDS advantage under corrupted inputs.
  • Because the method keeps only one historical feature and does not need future frames, it is compatible with streaming inference, though the dense BEV cost remains higher than sparse query-based methods.

Reading between the lines

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

  • Editorial inference: if the channel-difference motion signal is a learnable proxy for optical flow, the same alignment module could be extended with explicit per-object velocity supervision from LiDAR or radar, which the authors list as future work.
  • Editorial inference: the stop-gradient design of the consistency loss resembles teacher–student anchoring, so a temporal ensemble or exponential-moving-average version of the target heatmap might stabilize very long sequences even further.
  • Editorial inference: the same motion-guided alignment could be transferred to other dense BEV tasks, such as HD map construction or panoptic segmentation, where temporal consistency across moving objects also matters.
  • Editorial inference: the reported memory and accuracy numbers suggest that the performance ceiling of temporal fusion may be set more by alignment quality than by the number of frames, which would redirect future work toward better motion models rather than bigger memory banks.
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

2 major / 5 minor

Summary. The paper proposes OnlineBEV, a recurrent temporal fusion framework for multi-camera BEV-based 3D perception. Instead of storing and fusing a fixed-length window of historical BEV features, OnlineBEV keeps a single historical feature map and updates it recurrently. To handle motion-induced misalignment, the Motion-Guided BEV Fusion Network (MBFNet) extracts motion features as channel-wise differences between historical and current BEV features and uses them to guide deformable attention in Motion-Guided BEV Warping Attention (MGWA). A Heatmap-based Temporal Consistency Loss (HTC-loss) further encourages alignment by penalizing differences between heatmaps derived from the aligned historical features and the current features. Experiments on nuScenes report state-of-the-art 3D detection results (63.9% NDS on the test set), improved BEV segmentation and 3D occupancy prediction, robustness to synthetic corruptions, and generalization to Argoverse 2.

Significance. If the reported results are consistent and reproducible, the recurrent BEV fusion design with motion-guided alignment is a worthwhile contribution: it reduces memory usage relative to parallel temporal fusion, supports dense tasks such as BEV segmentation and occupancy prediction, and shows improved robustness under motion blur and occlusion. The paper also provides a useful complexity comparison and an ablation on Argoverse 2. However, the empirical support for the central state-of-the-art claim is weakened by an internal inconsistency between the main comparison table and the full-model ablation, by the absence of error bars or multiple runs given the very small test-set margin over SparseBEV, and by the lack of released code or checkpoints for independent verification. These issues are fixable but need to be addressed before the claims can be accepted.

major comments (2)
  1. [Section IV-C/IV-D, Tables I and IV] The full-model ablation and the main comparison table report different numbers for the same configuration. Table IV, Method (c), which is the complete OnlineBEV with recurrent fusion, MBFNet, and HTC-loss, lists 42.5 mAP / 51.9 NDS on the nuScenes val set with ResNet50 at 256x704. Table I lists the same configuration (ResNet50, 256x704, recurrent) as 44.4 mAP / 54.5 NDS. No difference in training schedule, pretraining, or augmentation is described in Section IV-B that would explain this gap. The discrepancy (1.9 mAP / 2.6 NDS) is larger than the individual gains attributed to MBFNet (1.3 mAP / 1.0 NDS) or HTC-loss (0.4 mAP / 0.5 NDS) in Table IV, making the component ablations and the headline comparison mutually inconsistent. The authors must either reconcile these numbers by reporting the missing details or rerun the affected experiments, and update the tables and claims accordingly.
  2. [Section IV-C, Table II] The test-set state-of-the-art claim rests on a margin of 0.2 mAP and 0.3 NDS over SparseBEV. The paper reports a single evaluation without error bars, multiple seeds, or any statistical significance test, and no code or checkpoints are provided. Given the size of the margin relative to typical run-to-run variation in this benchmark, the claim that OnlineBEV 'records state-of-the-art performance' is not robust unless the authors provide uncertainty estimates or otherwise demonstrate that the difference is consistent across runs.
minor comments (5)
  1. [Section III-A, Eq. (2)] Equation (2) uses the notation M(l)_{t-1}(p) for the motion feature, but the motion feature is defined in Eq. (1) as M(l)_t. Please make the time-index subscript consistent.
  2. [Section IV-D.2] The text says 'the baseline model is derived by disabling MFE in Method (b) of Table III,' but the ablation table is Table IV; Table III reports BEV segmentation and occupancy results. This appears to be a typo.
  3. [Table VII] The caption reads 'PERFORMANCE COMPARISON ON THE CORRUPTED NUSCENES TEST SET,' but the original inputs in the table (42.7 mAP / 53.4 NDS for SOLOFusion and 44.4 mAP / 54.5 NDS for OnlineBEV) match the validation set numbers in Table I. The caption should say 'validation set' unless the table reports test-set evaluations.
  4. [Section III-C] The consistency-loss weight w_cons is set to 2 'based on our experiments,' but no sensitivity analysis is reported. Since HTC-loss contributes only 0.5 NDS in the ablation, a small sweep or a fixed-choice justification would clarify how sensitive the improvement is to this hyperparameter.
  5. [Section III-A] The raw channel-wise difference in Eq. (1) is presented as a way to capture object motion, but the paper does not discuss cases where appearance changes (e.g., lighting, viewpoint, occlusion) also produce large differences. A brief discussion of why this simple difference is expected to encode displacement would strengthen the motivation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method's central claims are evaluated on an external benchmark with standard supervised losses, and no load-bearing step reduces to its own inputs.

full rationale

The paper's central empirical claim is benchmark performance on nuScenes, where val and test metrics are compared against externally published methods under a fixed public benchmark. The motion feature extractor (Eq. 1) and motion-guided warping attention (Eq. 2) are architectural components trained end-to-end with supervised detection losses; their contribution is assessed by ablations (Tables IV and V) on the held-out validation split, not by the objective they optimize. The HTC-loss (Eq. 6) is an auxiliary training regularizer: it directly minimizes heatmap disagreement, but its reported benefit is downstream mAP/NDS, an external measure, so the loss is not being presented as a prediction of that benefit. Setting w_cons to 2 is a normal validation-set hyperparameter choice, not a fitted value renamed as a prediction. The paper invokes no uniqueness theorem, and no load-bearing argument reduces to a self-citation chain. The only notable weakness is an internal inconsistency between the full-model ablation (Table IV Method (c): 42.5 mAP / 51.9 NDS) and the main comparison table (Table I OnlineBEV: 44.4 mAP / 54.5 NDS) under the same ResNet50/256x704 configuration; that is a reproducibility/correctness concern, not a circularity, because neither number is derived from the other by definition. Accordingly, no circular step is present and the honest finding is a score of 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a few domain assumptions about the BEV representation and the ability of attention to learn motion alignment. There are no invented physical entities. The only tuned free parameter is the consistency loss weight.

free parameters (1)
  • w_cons = 2.0
    The weight of the heatmap-based temporal consistency loss in Eq. 7, set to 2 based on experiments on the validation set.
assumptions (3)
  • domain assumption LSS-based depth estimation provides a reliable transformation from perspective view to BEV space.
    The framework builds on LSS; if depth estimates are poor, the BEV features are corrupted and the temporal fusion would not help.
  • domain assumption Deformable attention can learn to align BEV features using only motion features from feature differences.
    The MGWA module assumes the learned offsets can compensate for object motion in BEV space, without explicit optical flow or ground-truth motion.
  • domain assumption Ego-motion compensation accurately aligns static scene elements between frames.
    The method relies on ego-motion compensation for the static content, leaving only dynamic objects to be aligned by the network. Inaccurate compensation would break this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OnlineBEV: Recurrent Temporal Fusion in Bird's Eye View Representations for Multi-Camera 3D Perception." pith.science (2026). https://pith.science/paper/PNUWYHYN

@misc{pith2026250708644,
  author       = {Pith},
  title        = {Pith review of: OnlineBEV: Recurrent Temporal Fusion in Bird's Eye View Representations for Multi-Camera 3D Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PNUWYHYN}},
  note         = {Machine review of arXiv:2507.08644}
}
read the original abstract

Multi-view camera-based 3D perception can be conducted using bird's eye view (BEV) features obtained through perspective view-to-BEV transformations. Several studies have shown that the performance of these 3D perception methods can be further enhanced by combining sequential BEV features obtained from multiple camera frames. However, even after compensating for the ego-motion of an autonomous agent, the performance gain from temporal aggregation is limited when combining a large number of image frames. This limitation arises due to dynamic changes in BEV features over time caused by object motion. In this paper, we introduce a novel temporal 3D perception method called OnlineBEV, which combines BEV features over time using a recurrent structure. This structure increases the effective number of combined features with minimal memory usage. However, it is critical to spatially align the features over time to maintain strong performance. OnlineBEV employs the Motion-guided BEV Fusion Network (MBFNet) to achieve temporal feature alignment. MBFNet extracts motion features from consecutive BEV frames and dynamically aligns historical BEV features with current ones using these motion features. To enforce temporal feature alignment explicitly, we use Temporal Consistency Learning Loss, which captures discrepancies between historical and target BEV features. Experiments conducted on the nuScenes benchmark demonstrate that OnlineBEV achieves significant performance gains over the current best method, SOLOFusion. OnlineBEV achieves 63.9% NDS on the nuScenes test set, recording state-of-the-art performance in the camera-only 3D object detection task.

Figures

Figures reproduced from arXiv: 2507.08644 by the authors.

Figure 1
Figure 1. Different temporal fusion strategies. (a) Parallel temporal fusion aggregates historical BEV features within a fixed-length window at each time step. (b) Recurrent temporal fusion progressively updates historical BEV features over time. (c) Our OnlineBEV approach aligns historical BEV features with current BEV features to enable effective recurrent temporal fusion. Abstract—Multi-view camera-based 3D perception can … view at source ↗
Figure 2
Figure 2. The overall architecture of OnlineBEV. OnlineBEV aggregates historical BEV features with current BEV features using a recurrent structure. Before aggregation, MGWA aligns the historical BEV features to the current ones, guided by the motion features produced by MFE. During training, HTC-loss further facilitates the feature alignment process. Here, ‘sg’ denotes stop-gradient. experiences occlusion or motion blur. To … view at source ↗
Figure 3
Figure 3. Structure of Motion-Guided BEV Fusion Network (MBFNet). MBFNet comprises MFE and MGWA, where MFE generates motion features and MGWA applies deformable attention for feature alignment and fusion. further enhance alignment, we introduce the Heatmap-based Temporal Consistency Loss (HTC-loss), which encourages greater similarity between the heatmaps derived from Ht and Hˆ t. Finally, a detection head is applied to Ht to… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Examples of corrupted input images on nuScenes dataset. (a), (b), and (c) show the original input, motion blur, and occlusion, respectively. TABLE VI ABLATION STUDY CONDUCTED ON AGOVERSE 2 DATASET Method mAP ↑ CDS ↑ mATE ↓ mASE ↓ mAOE ↓ Baseline-S 13.6 8.2 0.967 0.452 …
Figure 5
Figure 5. Figure 5: Visualization of BEV features over time. Historical, aligned, and current BEV features are visualized over five consecutive time steps (t = 1 to t = 5) to show the temporal evolution and the effectiveness of the alignment process. TABLE VIII COMPUTATIONAL COMPLEXITY AN…
Figure 6
Figure 6. Figure 6: Qualitative results of SOLOFusion and OnlineBEV. Predicted 3D boxes are shown in red, while ground-truth 3D boxes are shown in green. This could be supervised using ground-truth motion data obtained from auxiliary sensors (e.g., LiDAR or radar). We plan to investigate …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 34 canonical work pages

  1. [1]

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

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

    arXiv 2021

  2. [2]

    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,” inProc. AAAI Conf. Artif. Intell., vol. 37, no. 2, 2023, pp. 1477–1485

    2023

  3. [3]

    Aedet: Azimuth-invariant multi-view 3d object detection,

    C. Feng, Z. Jie, Y . Zhong, X. Chu, and L. Ma, “Aedet: Azimuth-invariant multi-view 3d object detection,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2023, pp. 21 580–21 588

    work page 2023

  4. [4]

    Fb- bev: Bev representation from forward-backward view transformations,

    Z. Li, Z. Yu, W. Wang, A. Anandkumar, T. Lu, and J. M. Alvarez, “Fb- bev: Bev representation from forward-backward view transformations,” inProc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 6919–6928

    work page 2023

  5. [5]

    Sa-bev: Generating semantic- aware bird’s-eye-view feature for multi-view 3d object detection,

    J. Zhang, Y . Zhang, Q. Liu, and Y . Wang, “Sa-bev: Generating semantic- aware bird’s-eye-view feature for multi-view 3d object detection,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 3348–3357

    work page 2023

  6. [6]

    Detr3d: 3d object detection from multi-view images via 3d-to-2d queries,

    Y . Wang, V . C. Guizilini, T. Zhang, Y . Wang, H. Zhao, and J. Solomon, “Detr3d: 3d object detection from multi-view images via 3d-to-2d queries,” inProc. Conf. Robot Learn.PMLR, 2022, pp. 180–191

    work page 2022

  7. [7]

    Petr: Position embedding transformation for multi-view 3d object detection,

    Y . Liu, T. Wang, X. Zhang, and J. Sun, “Petr: Position embedding transformation for multi-view 3d object detection,” inProc. Eur . Conf. Comput. Vis.Springer, 2022, pp. 531–548

    work page 2022

  8. [8]

    Viewpoint equivariance for multi-view 3d object detection,

    D. Chen, J. Li, V . Guizilini, R. A. Ambrus, and A. Gaidon, “Viewpoint equivariance for multi-view 3d object detection,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2023, pp. 9213–9222

    work page 2023

Show all 43 references
  1. [9]

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

    J. Philion and S. Fidler, “Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,” inProc. Eur . Conf. Comput. Vis.Springer, 2020, pp. 194–210

    2020

  2. [10]

    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,”arXiv:2203.17054, 2022

    2022 arXiv

  3. [11]

    Bevstereo: Enhancing depth estimation in multi-view 3d object detection with 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 temporal stereo,” inProc. AAAI Conf. Artif. Intell., vol. 37, no. 2, 2023, pp. 1486–1494

    2023

  4. [12]

    Predict to detect: Prediction- guided 3d object detection using sequential images,

    S. Kim, Y . Kim, I.-J. Lee, and D. Kum, “Predict to detect: Prediction- guided 3d object detection using sequential images,” inProc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 18 057–18 066

    2023

  5. [13]

    Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection,

    J. Park, C. Xu, S. Yang, K. Keutzer, K. M. Kitani, M. Tomizuka, and W. Zhan, “Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection,” inInt. Conf. Learn. Representations, 2022

    2022

  6. [14]

    Sparsebev: High- performance sparse 3d object detection from multi-camera videos,

    H. Liu, Y . Teng, T. Lu, H. Wang, and L. Wang, “Sparsebev: High- performance sparse 3d object detection from multi-camera videos,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 18 580–18 590

    2023

  7. [15]

    Exploring object-centric temporal modeling for efficient multi-view 3d object detection,

    S. Wang, Y . Liu, T. Wang, Y . Li, and X. Zhang, “Exploring object-centric temporal modeling for efficient multi-view 3d object detection,” inProc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 3621–3631

    2023

  8. [16]

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

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

    2010 arXiv

  9. [17]

    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,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2020, pp. 11 621–11 631

    2020

  10. [18]

    One stage monocular 3d object detection utilizing discrete depth and orientation representation,

    M. A. Haq, S.-J. Ruan, M.-E. Shao, Q. M. U. Haq, P.-J. Liang, and D.-Q. Gao, “One stage monocular 3d object detection utilizing discrete depth and orientation representation,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 11, pp. 21 630–21 640, 2022

    2022

  11. [19]

    Occlusion-aware plane-constraints for monocular 3d object detection,

    H. Yao, J. Chen, Z. Wang, X. Wang, P. Han, X. Chai, and Y . Qiu, “Occlusion-aware plane-constraints for monocular 3d object detection,” IEEE Trans. Intell. Transp. Syst., 2023

    2023

  12. [20]

    Shape-aware monocular 3d object detection,

    W. Chen, J. Zhao, W.-L. Zhao, and S.-Y . Wu, “Shape-aware monocular 3d object detection,”IEEE Trans. Intell. Transp. Syst., vol. 24, no. 6, pp. 6416–6424, 2023

    2023

  13. [21]

    Bevformer: Learning bird’s-eye-view representation from multi-camera JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, JULY 2025 10 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 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, JULY 2025 10 images via spatiotemporal transformers,” inProc. Eur . Conf. Comput. Vis.Spring...

    2025

  14. [22]

    A stereo matching algorithm with an adaptive window: Theory and experiment,

    T. Kanade and M. Okutomi, “A stereo matching algorithm with an adaptive window: Theory and experiment,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 16, no. 9, pp. 920–932, 1994

    1994

  15. [23]

    Detr4d: Direct multi-view 3d object detection with sparse attention,

    Z. Luo, C. Zhou, G. Zhang, and S. Lu, “Detr4d: Direct multi-view 3d object detection with sparse attention,”arXiv:2212.07849, 2022

    2022 arXiv

  16. [24]

    Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion,

    X. Lin, T. Lin, Z. Pei, L. Huang, and Z. Su, “Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion,”arXiv:2211.10581, 2022

    2022 arXiv

  17. [25]

    Consistency-based semi- supervised learning for object detection,

    J. Jeong, S. Lee, J. Kim, and N. Kwak, “Consistency-based semi- supervised learning for object detection,” vol. 32, pp. 10 759–10 768, 2019

    2019

  18. [26]

    Semi-supervised semantic segmenta- tion with cross-consistency training,

    Y . Ouali, C. Hudelot, and M. Tami, “Semi-supervised semantic segmenta- tion with cross-consistency training,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2020, pp. 12 674–12 684

    2020

  19. [27]

    An unsupervised temporal consistency (tc) loss to improve the performance of semantic segmentation networks,

    S. Varghese, S. Gujamagadi, M. Klingner, N. Kapoor, A. Bar, J. D. Schneider, K. Maag, P. Schlicht, F. Huger, and T. Fingscheidt, “An unsupervised temporal consistency (tc) loss to improve the performance of semantic segmentation networks,” inProc. IEEE/CVF Conf. Comput. Vis. P...

    2021

  20. [28]

    Mask- free video instance segmentation,

    L. Ke, M. Danelljan, H. Ding, Y .-W. Tai, C.-K. Tang, and F. Yu, “Mask- free video instance segmentation,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2023, pp. 22 857–22 866

    2023

  21. [29]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2018, pp. 7132–7141

    2018

  22. [30]

    Dropout: a simple way to prevent neural networks from overfitting,

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdi- nov, “Dropout: a simple way to prevent neural networks from overfitting,” J. Mach. Learn. Res., vol. 15, no. 1, pp. 1929–1958, 2014

    1929

  23. [31]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv:1607.06450, 2016

    2016 arXiv

  24. [32]

    Center-based 3d object detection and tracking,

    T. Yin, X. Zhou, and P. Krahenbuhl, “Center-based 3d object detection and tracking,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2021, pp. 11 784–11 793

    2021

  25. [33]

    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,” inProc. IEEE/CVF Int. Conf. Comput. Vis., 2017, pp. 2980–2988

    2017

  26. [34]

    Petrv2: A unified framework for 3d perception from multi-camera images,

    Y . Liu, J. Yan, F. Jia, S. Li, A. Gao, T. Wang, and X. Zhang, “Petrv2: A unified framework for 3d perception from multi-camera images,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 3262–3272

    2023

  27. [35]

    Argoverse 2: Next generation datasets for self-driving perception and forecasting,

    B. Wilson, W. Qi, T. Agarwal, J. Lambert, J. Singh, S. Khandelwal, B. Pan, R. Kumar, A. Hartnett, J. K. Ponteset al., “Argoverse 2: Next generation datasets for self-driving perception and forecasting,”arXiv preprint arXiv:2301.00493, 2023

    2023 arXiv

  28. [36]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2016, pp. 770–778

    2016

  29. [37]

    An energy and gpu- computation efficient backbone network for real-time object detection,

    Y . Lee, J.-w. Hwang, S. Lee, Y . Bae, and J. Park, “An energy and gpu- computation efficient backbone network for real-time object detection,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops, 2019, pp. 0–0

    2019

  30. [38]

    Decoupled weight decay regularization,

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

    2019

  31. [39]

    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?” inProc. IEEE/CVF Int. Conf. Comput. Vis., 2021, pp. 3142–3152

    2021

  32. [40]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2009, pp. 248–255

    2009

  33. [41]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022, pp. 11 976–11 986

    2022

  34. [42]

    Benchmarking robustness of 3d object detection to common corruptions,

    Y . Dong, C. Kang, J. Zhang, Z. Zhu, Y . Wang, X. Yang, H. Su, X. Wei, and J. Zhu, “Benchmarking robustness of 3d object detection to common corruptions,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 1022–1032

    2023

  35. [43]

    Benchmarking the robustness of lidar-camera fusion for 3d object detection,

    K. Yu, T. Tao, H. Xie, Z. Lin, T. Liang, B. Wang, P. Chen, D. Hao, Y . Wang, and X. Liang, “Benchmarking the robustness of lidar-camera fusion for 3d object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3188–3198. BI...

    2023

Pith tools

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