Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Refine-and-Contrast: Adaptive Instance-Aware BEV Representations for Multi-UAV Collaborative Object Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a multi-UAV 3D detector can skip uniform bird's-eye-view computation and refine only foreground grid cells, reaching near-upper-bound accuracy at a fraction of the compute.

desk verdict A clean, incremental BEVFormer extension that recovers part of the high-resolution accuracy at low resolution; the main thing to check is the train/test mask mismatch in its refinement module. read the letter →

arxiv 2508.12684 v1 pith:KUABRKFK submitted 2025-08-18 cs.CV

Zhongyao Li , Peirui Cheng , Liangjin Zhao , Chen Chen , Yundu Li , Zhechao Wang , Xue Yang , Xian Sun
show 1 more author
Zhirui Wang
This is my paper
classification cs.CV
keywords multi-UAVperceptioncollaborative3Ddetectionbird's-eye-viewrepresentationinstance-awarerefinementcontrastivelearningBEVspatialcross-attentioncomputation-accuracytrade-off
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 treating every bird's-eye-view (BEV) grid cell as equally important is wasteful for multi-UAV 3D detection, where objects are small, sparse, and clustered. It introduces AdaBEV, which marks BEV cells that fall inside 2D detection boxes as foreground, re-runs spatial cross-attention on only those cells at 4x finer subdivisions, and applies an InfoNCE contrastive loss to separate instance features from background features. On the Air-Co-Pred benchmark, AdaBEV at 50x50 BEV raises AP from 0.759 to 0.783 over the BEVFormer baseline while adding just 0.07 GFLOPs, and across model scales it recovers more than half of the gain that jumping to 200x200 BEV would give. If correct, this means near-upper-bound accuracy is achievable at low BEV resolution, which matters directly for the limited compute budgets of drones.

What carries the argument

The load-bearing mechanism is a foreground mask derived from 2D detection boxes, used to gate where extra computation happens. BG-RM refines only masked BEV cells through a 4x4 spatial subdivision of the standard spatial cross-attention, and IBCL enforces foreground-background separation in BEV space with a contrastive loss. The mask is the object that transfers 2D semantic knowledge into the BEV plane and is what makes the refinement 'instance-aware'.

What would settle it

Run AdaBEV at inference twice on the same validation set: once with the mask built from the perspective-aware branch's predicted boxes and once with the mask built from ground-truth boxes, then compare AP. If the predicted-box mask preserves most of the 0.759-to-0.783 gain, the refinement mechanism is robust; if the gain collapses toward the baseline, the reported improvement depends on an unrealistic mask. Additionally, report precision and recall of the predicted foreground mask against the ground-truth mask.

Watch

Extended reading notes

Core claim

AdaBEV's central discovery is that BEV representation learning does not have to be spatially uniform. The Box-Guided Refinement Module (BG-RM) projects predicted 2D boxes into BEV space to build a foreground mask, subdivides each masked grid cell into a 4x4 sub-grid, computes spatial cross-attention at that finer resolution, then averages the sub-grid outputs back into the coarse cell while leaving background cells untouched. The Instance-Background Contrastive Learning (IBCL) module then crops BEV features inside ground-truth 3D boxes as instance features, samples background patches at a minimum distance, and applies an InfoNCE loss to push the two apart. Together these two modules produce

Load-bearing premise

The foreground mask produced at inference time by the perspective-aware 2D detector is accurate enough that refining only those BEV cells and ignoring everything else does not drop objects; the paper trains the mask on ground-truth boxes but never measures how well the predicted boxes reproduce that mask.

Editorial extensions

If this is right

  • At 50x50 BEV resolution, AdaBEV-tiny reaches 0.775 AP versus the 0.786 upper bound obtained with 200x200 BEV, while using 141.56 GFLOPs instead of 364.06 GFLOPs.
  • On a ResNet-50 backbone, AdaBEV improves AP from 0.759 to 0.783 over the BEVFormer baseline and also reduces translation, scale, and orientation errors.
  • The gains persist at higher capacity: with ResNet-101 at 150x150 BEV, AdaBEV reaches 0.910 AP versus 0.904 for the baseline.
  • Scaling collaboration from 2 to 4 UAVs gives a large improvement in AdaBEV (0.595 to 0.783 AP), with diagonal viewpoint pairs outperforming adjacent pairs.
  • The ablation shows each component contributes: PAS alone gives 0.764 AP, adding BG-RM gives 0.775, and adding IBCL reaches 0.783.

Reading between the lines

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

  • If the refine-and-contrast mechanism is general, it should transfer to other BEV detectors beyond the BEVFormer family, such as BEVDet or BEVDepth, by grafting the same foreground-mask refinement onto their spatial aggregation; this is a testable extension the paper does not run.
  • The contrastive loss design is not fully ablated: the number of background samples, the minimum distance threshold, and the temperature are fixed, and a simpler margin-based or focal loss might match IBCL at lower tuning cost.
  • Because BG-RM's mask is trained on ground-truth boxes but applied at inference to predicted boxes, part of the reported gain could come from a train/inference mask mismatch; measuring mask precision and recall at test time would show how much of the improvement survives with imperfect masks.
  • The foreground mask naturally defines a sparse set of BEV cells that matter, so a multi-UAV communication protocol could transmit only foreground BEV features between drones, potentially reducing bandwidth further than the paper explores.
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

3 major / 5 minor

Summary. The paper proposes AdaBEV, a BEVFormer-based 3D detection framework for multi-UAV collaborative perception. Two modules are added: a Box-Guided Refinement Module (BG-RM) that refines only BEV grids whose projected 3D reference points fall inside 2D detection boxes, and an Instance-Background Contrastive Learning (IBCL) module that contrasts instance features extracted from ground-truth 3D boxes against randomly sampled background features. Experiments on the Air-Co-Pred dataset report consistent AP improvements over the BEVFormer baseline at low BEV resolution with negligible GFLOPs overhead, e.g., 0.783 vs. 0.759 at 50x50 resolution with ResNet-50, and claim that these gains approach the high-resolution upper bound.

Significance. The core idea is timely and practically motivated: in aerial images objects are small and sparse, so uniform BEV refinement wastes computation and may dilute foreground features. The combination of 2D-box-guided refinement and contrastive separation is a plausible, low-overhead way to obtain instance-aware BEV representations. If the reported gains are robust, the method is a useful building block for resource-constrained collaborative UAV perception. The paper also provides a useful comparison across model scales and an ablation of the two proposed modules. However, the central empirical claim depends on a training/inference mismatch in the refinement mask, on small single-run accuracy differences, and on efficiency statements that are only partially supported by the paper's own tables.

major comments (3)
  1. [Section 3.2, Eq. (3), and Table 4] The training-time foreground mask is built from ground-truth boxes: "During training, we replace the predicted boxes with ground-truth annotations" (Section 3.2). At inference the mask is built from boxes produced by the perspective-aware supervision (PAS) branch. The paper never measures the quality of the PAS boxes, the precision/recall of the resulting mask, or the effect of using predicted boxes during training. Since BG-RM contributes only +0.011 AP over PAS alone (0.764 to 0.775 in Table 4), the train/test mask mismatch is load-bearing for the claimed benefit. Please add an ablation that trains BG-RM with predicted-box masks (or a mix of predicted and GT masks), and report PAS box AP and mask IoU/precision/recall on the val set.
  2. [Section 4.3 and Table 2] The efficiency claims are internally inconsistent. The text says increasing BEV resolution from 50x50 to 200x200 "leads to about 2.5x increases in GFLOPs across model scales," but Table 2 gives ratios of 364.06/141.49 = 2.57 (tiny), 608.88/386.31 = 1.58 (small), and 1266.82/880.67 = 1.44 (base). The subsequent claim that the method "achieves over 50% of the AP improvement" of high resolution is also true only for the tiny model: for small the improvement is 0.004/0.014 = 29%, and for base it is 0.004/0.010 = 40%. The accuracy-computation claim should be stated separately per model scale. Note also that the parameter overhead is not negligible in Table 2 (41.01 vs. 33.57M, +22%), so "negligible overhead" should be restricted to GFLOPs, not parameters.
  3. [Section 4, Tables 1, 2, and 4] All quantitative results appear to be single runs with no error bars, standard deviations, or statistical significance tests. Several of the key comparisons are small in absolute terms, e.g., AdaBEV-R vs. BEVFormer-small lower bound: 0.902 vs. 0.898; AdaBEV-R vs. BEVFormer-base lower bound: 0.934 vs. 0.930. Without repeated seeds, these differences could be within run-to-run variation, which weakens the central accuracy-computation trade-off claim. Please report mean +/- std over at least three seeds for the main comparisons in Tables 2 and 4, or provide a significance test.
minor comments (5)
  1. [Section 3.3 and Figure 2 caption] IBCL is described as "self-supervised" in the Figure 2 caption, but the instance extractor uses ground-truth 3D boxes to define instance features and positive pairs. This is supervised by labels, not self-supervised. Please revise the wording.
  2. [Section 4.3] The sentence "In general, our method achieves over 50% of the AP improvement ... with less than 1.5% additional computation" should be corrected per model scale, since the 50% figure holds only for the tiny model and the small/base models fall below it.
  3. [Figures 2 and 3] The figures use labels such as "Fore grid", "Back grid", and "IB-Contrastive Learning" that are not formally defined in the text. Please align figure terminology with the notation in Section 3 (e.g., B_f, B_b, IBCL).
  4. [Figure 5] The rightmost panel is labeled "LIDAR_TOP (GT)", but it is unclear whether this is a LiDAR point cloud, a ground-truth BEV occupancy map, or ground-truth boxes. Please clarify the visualization source.
  5. [Table 1] BEVLongTerm and BEVDet4D-Stereo are both cited as "Huang and Huang 2022" in the caption/references. If these are different methods, they need distinct citations; if not, the table should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; BG-RM and IBCL use ground-truth supervision at training time but the reported detections are produced by an independent decoder, so the central claim does not reduce to its inputs.

full rationale

The paper's chain is an empirical architecture-plus-loss study, not a derivation. BG-RM (Section 3.2) constructs a foreground mask from 2D boxes: at training these are GT annotations ('During training, we replace the predicted boxes with ground-truth annotations for more accurate supervision and stable optimization'), while at inference they come from the PAS branch. This is a train/inference supervision mismatch and a robustness caveat, but it is not circular: the mask only selects which BEV grids receive a 4x4 refined Spatial Cross-Attention; the final 3D detections are produced by the BEVFormer DETR decoder from the refined BEV feature and are not defined by the mask or the PAS boxes. IBCL (Section 3.3) uses projected GT 3D boxes only to form a contrastive loss (Eq. 11) during training; it is not part of the inference graph, so the reported AP (Tables 1 and 2) is not computed from the labels used to supervise the loss. There is no load-bearing self-citation: the Air-Co-Pred dataset (Wang et al. 2024) and BEVFormer baseline are standard shared prior work, not uniqueness theorems or ansatz-bearing citations. The improvement claims are benchmark comparisons against external methods with matched GFLOPs and resolution; the ablation (Table 4) shows incremental gains from PAS, BG-RM, and IBCL that are not forced by construction. The predicted-box/GT-box mask mismatch is a legitimate correctness risk (PAS 2D box quality and mask precision/recall are unmeasured at inference), but it does not make the result equivalent to its input by definition, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper relies on a standard BEVFormer backbone, deformable attention, and InfoNCE loss (standard math), plus several domain assumptions about the Air-Co-Pred simulation and the reliability of box-guided foreground masks. The main free parameters (loss weights, subdivision factor, sampling distance) are hand-set and mostly unablated. No new physical entities are introduced.

free parameters (6)
  • IBCL loss weight lambda_2 = 2.0
    Set by hand in Eq. 12; no sensitivity analysis reported.
  • PAS loss weight lambda_1 = 1.0
    Set by hand in Eq. 12; no sensitivity analysis reported.
  • BG-RM subdivision factor r = 4
    Default in Section 3.2; no ablation over r is provided.
  • background sampling minimum distance d_min = 4 meters
    Chosen in Section 4.1 to avoid semantic contamination; no sensitivity analysis.
  • number of background samples = 200
    Fixed in Section 4.1; no ablation is reported.
  • InfoNCE temperature tau = not specified
    Appears in Eq. 11 but no value is given anywhere in the paper.
assumptions (5)
  • standard math Deformable attention as in Deformable DETR and BEVFormer is a valid feature aggregation operator
    Used in SCA and BG-RM; cited from Zhu et al. 2020b.
  • domain assumption InfoNCE contrastive loss encourages instance-background separability in BEV space
    Assumed in IBCL; the objective may be sensitive to sampling and temperature.
  • domain assumption Foreground BEV grids can be identified by projecting pillar samples into 2D boxes
    The validity depends on calibration and the accuracy of the boxes; see Eq. 1-3.
  • domain assumption Air-Co-Pred simulated data is representative of real multi-UAV perception
    All experiments are on a simulation; no real-world validation is provided.
  • ad hoc to paper Ground-truth boxes at training for mask generation do not distort the measured contribution of BG-RM
    The paper trains with GT boxes and infers with predicted boxes without evaluating this mismatch.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Refine-and-Contrast: Adaptive Instance-Aware BEV Representations for Multi-UAV Collaborative Object Detection." pith.science (2026). https://pith.science/paper/KUABRKFK

@misc{pith2026250812684,
  author       = {Pith},
  title        = {Pith review of: Refine-and-Contrast: Adaptive Instance-Aware BEV Representations for Multi-UAV Collaborative Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KUABRKFK}},
  note         = {Machine review of arXiv:2508.12684}
}
read the original abstract

Multi-UAV collaborative 3D detection enables accurate and robust perception by fusing multi-view observations from aerial platforms, offering significant advantages in coverage and occlusion handling, while posing new challenges for computation on resource-constrained UAV platforms. In this paper, we present AdaBEV, a novel framework that learns adaptive instance-aware BEV representations through a refine-and-contrast paradigm. Unlike existing methods that treat all BEV grids equally, AdaBEV introduces a Box-Guided Refinement Module (BG-RM) and an Instance-Background Contrastive Learning (IBCL) to enhance semantic awareness and feature discriminability. BG-RM refines only BEV grids associated with foreground instances using 2D supervision and spatial subdivision, while IBCL promotes stronger separation between foreground and background features via contrastive learning in BEV space. Extensive experiments on the Air-Co-Pred dataset demonstrate that AdaBEV achieves superior accuracy-computation trade-offs across model scales, outperforming other state-of-the-art methods at low resolutions and approaching upper bound performance while maintaining low-resolution BEV inputs and negligible overhead.

Figures

Figures reproduced from arXiv: 2508.12684 by the authors.

Figure 1
Figure 1. Comparison between our and previous meth￾ods. We break the conventional paradigm of uniform BEV modeling by adaptively refining target regions to produce instance-aware BEV representations. Each grid represents a BEV query. a widely adopted representation in multi-agent aerial sys￾tems. Current mainstream BEV-based 3D detection methods predominantly follow a Transformer-based paradigm (Li et al. 2022; Pan et al. 202… view at source ↗
Figure 2
Figure 2. Overall architecture of AdaBEV. AdaBEV consists of two key components: a box-guided refinement module (BG￾RM) that enhances foreground features using detections from a perspective-aware supervision (PAS) branch, and an instance￾background contrastive learning (IBCL) module that enforces discriminative BEV representations via self-supervised learning. During inference, multi-view RGB images are first fed into an imag… view at source ↗
Figure 4
Figure 4. Illustration of Instance-Background Con￾trastive Learning (IBCL). The instance extractor (IE) projects ground-truth 3D boxes onto BEV to extract instance features, while the background sampler (BS) samples fea￾tures from background regions. IBCL employs an InfoNCE loss to enforce instance-background feature separability. overfitting and overlook the intrinsic structure of BEV fea￾tures. To address this, IBCL introdu… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Visualization of BEV feature maps in BEVFormer and AdaBEV. Compared to the BEVFormer, AdaBEV produces more detailed and semantically discriminative BEV features in instance regions. The highlighted regions show clearer object boundaries and improved separability from b…
Figure 6
Figure 6. Figure 6: Visualization results of AdaBEV on the Air-Co-Pred val set. We show the 3D bboxes predictions in multi-UAV images and the bird’s-eye view. ble 3, increasing the number of collaborative UAVs leads to consistent improvements across all metrics. In particu￾lar, using four…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 20 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    H.; Vora, S.; Liong, V

    Caesar, H.; Bankiti, V.; Lang, A. H.; Vora, S.; Liong, V. E.; Xu, Q.; Krishnan, A.; Pan, Y.; Baldan, G.; and Beijbom, O. 2020. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11621--11631

    2020

  4. [4]

    Chang, X.; Yang, C.; Wu, J.; Shi, X.; and Shi, Z. 2018. A surveillance system for drone localization and tracking using acoustic arrays. In 2018 IEEE 10th Sensor Array and Multichannel Signal Processing Workshop (SAM), 573--577. IEEE

    work page 2018

  5. [5]

    Chen, T.; Kornblith, S.; Norouzi, M.; and Hinton, G. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, 1597--1607. PmLR

    2020

  6. [6]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 770--778

    2016

  7. [7]

    Hu, Y.; Fang, S.; Lei, Z.; Zhong, Y.; and Chen, S. 2022. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35: 4874--4886

    work page 2022

  8. [8]

    Huang, J.; and Huang, G. 2022. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054

    arXiv 2022

Show all 32 references
  1. [9]

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

    2021 arXiv

  2. [10]

    Jiang, X.; Li, S.; Liu, Y.; Wang, S.; Jia, F.; Wang, T.; Han, L.; and Zhang, X. 2024. Far3d: Expanding the horizon for surround-view 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, volume 38, 2561--2569

    2024

  3. [11]

    Li, Y.; Ge, Z.; Yu, G.; Yang, J.; Wang, Z.; Shi, Y.; Sun, J.; and Li, Z. 2023 a . Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 1477--1485

    2023

  4. [12]

    Li, Z.; Wang, W.; Li, H.; Xie, E.; Sima, C.; Lu, T.; Qiao, Y.; and Dai, J. 2022. BEVFormer: Learning Bird’s-Eye-View Representation from Multi-camera Images via Spatiotemporal Transformers. In European Conference on Computer Vision, 1--18

    2022

  5. [13]

    Li, Z.; Yu, Z.; Wang, W.; Anandkumar, A.; Lu, T.; and Alvarez, J. M. 2023 b . Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 6919--6928

    2023

  6. [14]

    Lin, T.-Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Doll \'a r, P.; and Zitnick, C. L. 2014. Microsoft coco: Common objects in context. In European conference on computer vision, 740--755. Springer

    2014

  7. [15]

    Liu, Y.; Wang, T.; Zhang, X.; and Sun, J. 2022. Petr: Position embedding transformation for multi-view 3d object detection. In European conference on computer vision, 531--548. Springer

    2022

  8. [16]

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

    2017 arXiv

  9. [17]

    Nguyen, H. P. D.; and Nguyen, D. D. 2021. Drone application in smart cities: The general overview of security vulnerabilities and countermeasures for data communication. Development and Future of Internet of Drones (IoD): Insights, Trends and Road Ahead, 185--210

    2021

  10. [18]

    Oord, A. v. d.; Li, Y.; and Vinyals, O. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748

    2018 arXiv

  11. [19]

    Pan, C.; Yaman, B.; Velipasalar, S.; and Ren, L. 2024. Clip-bevformer: Enhancing multi-view image-based bev detector with ground truth flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15216--15225

    2024

  12. [20]

    Park, D.; Ambrus, R.; Guizilini, V.; Li, J.; and Gaidon, A. 2021. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF international conference on computer vision, 3142--3152

    2021

  13. [21]

    Philion, J.; and Fidler, S. 2020. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European conference on computer vision, 194--210. Springer

    2020

  14. [22]

    B.; Singh, R.; Calyam, P.; and Das, S

    Qu, C.; Sorbelli, F. B.; Singh, R.; Calyam, P.; and Das, S. K. 2023. Environmentally-aware and energy-efficient multi-drone coordination and networking for disaster response. IEEE transactions on network and service management, 20(2): 1093--1109

    2023

  15. [23]

    Tian, P.; Wang, Z.; Cheng, P.; Wang, Y.; Wang, Z.; Zhao, L.; Yan, M.; Yang, X.; and Sun, X. 2024. Ucdnet: Multi-uav collaborative 3d object detection network by reliable feature mapping. IEEE Transactions on Geoscience and Remote Sensing

    2024

  16. [24]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  17. [25]

    Wang, S.; Jiang, X.; and Li, Y. 2023. Focal-petr: Embracing foreground for efficient multi-camera 3d object detection. IEEE Transactions on Intelligent Vehicles, 9(1): 1481--1489

    2023

  18. [26]

    Wang, Z.; Cheng, P.; Chen, M.; Tian, P.; Wang, Z.; Li, X.; Yang, X.; and Sun, X. 2024. Drones help drones: A collaborative framework for multi-drone object trajectory prediction and beyond. Advances in Neural Information Processing Systems, 37: 64604--64628

    2024

  19. [27]

    Wang, Z.; Huang, Z.; Fu, J.; Wang, N.; and Liu, S. 2023. Object as query: Lifting any 2d object detector to 3d detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 3791--3800

    2023

  20. [28]

    Yang, C.; Chen, Y.; Tian, H.; Tao, C.; Zhu, X.; Zhang, Z.; Huang, G.; Li, H.; Qiao, Y.; Lu, L.; et al. 2023. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF conference on computer vision a...

    2023

  21. [29]

    Yang, C.; Lin, T.; Huang, L.; and Crowley, E. J. 2024. Widthformer: Toward efficient transformer-based bev view transformation. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 8457--8464. IEEE

    2024

  22. [30]

    a henb \

    Zhou, B.; and Kr \"a henb \"u hl, P. 2022. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 13760--13769

    2022

  23. [31]

    Zhu, P.; Zheng, J.; Du, D.; Wen, L.; Sun, Y.; and Hu, Q. 2020 a . Multi-drone-based single object tracking with agent sharing network. IEEE Transactions on Circuits and Systems for Video Technology, 31(10): 4058--4070

    2020

  24. [32]

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

    2020 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.