Pith. sign in

REVIEW 4 major objections 4 minor 28 references

CoGoal3D claims that treating collaborative V2X fusion as a 3D alignment problem rather than a 2D BEV warp lifts 3D detection AP by roughly 10 points on three real-world datasets.

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 ↗

CoGoal3D, a broadcast two-stage collaborative 3D detector with 3D-aware fusion and point-reconstruction refinement, reports state-of-the-art 3D AP on three real-world V2X datasets.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A well-engineered two-stage collaborative 3D detector whose internal ablations support each module, but whose headline 10-point SOTA margins rest on baseline comparisons that are not apples-to-apples. the 4 major comments →

arxiv 2607.19036 v1 pith:PC3L3PML submitted 2026-07-21 cs.CV cs.AI

CoGoal3D: Collaborative 3D Object Detection with 3D-Aware Fusion and Refinement

Zhihao Yang , Zhiyu Xiang , Peng Xu , Tianyu Pu , Kai Wang , Eryun Liu , Dongping Zhang , Yong Ding This is my paper
classification cs.CV cs.AI
keywords Autonomous drivingCollaborative perception3D object detectionV2XLiDARBird's-eye view fusionData augmentationPoint cloud reconstruction
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

CoGoal3D is a two-stage collaborative 3D object detector that treats the alignment of multiple agents' LiDAR features as a 3D problem rather than a 2D BEV warping problem. The paper argues that ignoring height and attitude differences between collaborating vehicles or roadside units is a major cause of degraded 3D detection in broadcast-based V2X systems, and that correcting it with 3D position encoding and deformable cross attention, then refining proposals with an auxiliary point-reconstruction task, produces large gains. Across DAIR-V2X, V2V4Real, and V2X-Real, the method reports 3D AP@0.7 improvements of 10.86%, 10.34%, and 10.18% over prior state-of-the-art methods, and the full pipeline runs at 16.8 FPS. This matters because broadcast-based collaboration is the low-latency communication paradigm for practical V2X; if the gains hold, collaborative 3D detection can approach handshake-based accuracy without extra communication rounds.

Core claim

The authors' central claim is that the spatial misalignment caused by different sensor mounting heights and attitudes among collaborators is not adequately handled by standard 2D BEV feature warping, and that explicit 3D-aware fusion and refinement recovers the lost accuracy. In stage one, the multiscale 3D-Aware Global Fusion module computes relative pose transforms, encodes 3D pillar-center positions, and uses deformable cross attention with 3D-aware reference points to jointly align and fuse features. In stage two, the Reconstruction-Guided Local Refinement module reconstructs the 3D points inside each proposal under supervision from a Ground-Truth Optimization that re-aligns duplicated p

What carries the argument

The load-bearing mechanism is the multiscale 3D-Aware Global Fusion (3D-AGF) module: it augments BEV features with 3D position encoding derived from relative pose transforms and performs deformable cross attention using 3D-aware reference points, so that alignment and feature aggregation happen jointly rather than as a separate 2D warp. The second-stage Reconstruction-Guided Local Refinement (RGLR) then adds fine-grained 3D supervision through an auxiliary point-reconstruction head, guided by an optimized ground truth (GTO), which sharpens the estimated 3D boxes.

Load-bearing premise

The reported 10-point AP gains assume the comparison numbers for prior methods were produced under the same backbone, detection range, ground-truth annotations, and augmentation protocol as CoGoal3D; the paper does not spell out exactly which baselines were retrained and with which settings.

What would settle it

Re-train all baseline methods in the same codebase with identical PointPillars backbone, the same detection range and GT annotations, and the same augmentation settings, then re-evaluate on the three datasets; if the ~10-point margins collapse, the central generalisation claim fails. A quicker internal check: remove the 3D position encoding from the fusion module alone; Table 4 predicts a 3.85 AP@0.7 drop on DAIR-V2X, so a run that shows no such drop would contradict the claimed mechanism.

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

If this is right

  • Broadcast-based collaborative perception, which requires only one communication round, can reach or exceed the 3D detection accuracy of handshake-based methods that were previously considered more accurate.
  • The reconstruction-guided refinement can be added to existing broadcast fusion detectors as a second stage, potentially improving their 3D geometry without changing the communication protocol.
  • The multi-agent collaborative data augmentation strategy improves not only CoGoal3D but also other published collaborative detectors, as the paper's Table 6 shows.
  • The method retains robustness under injected pose noise and transmission latency, suggesting it can operate under imperfect real-world V2X conditions.

Where Pith is reading between the lines

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

  • If the reported margins hold, they imply that height and attitude discrepancies, not just localization noise, are a primary error source in real-world V2X 3D detection; methods that ignore 3D alignment are leaving a large, systematic gap on the table.
  • The Ground Truth Optimization idea—using duplicated observations of the same object from multiple agents to derive a better supervision signal—could transfer to other multi-sensor tasks such as cooperative mapping, multi-view reconstruction, or sensor calibration.
  • The local-rotation data augmentation might become the default for large-scale V2X: as the number of agents grows and detection ranges become asymmetric, information loss from global augmentations compounds, so a decentralized augmentation scheme could generalize better.
  • A testable extension would be applying the 3D-aware fusion module to cooperative prediction or planning features, not just object detection, where misalignment could corrupt downstream motion forecasting.
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 / 4 minor

Summary. CoGoal3D is a two-stage broadcast-based collaborative 3D object detection framework. In stage 1, a multiscale 3D-Aware Global Fusion (3D-AGF) module uses 3D position encoding and deformable cross-attention to align and fuse BEV features from collaborating agents. In stage 2, a Reconstruction-Guided Local Refinement (RGLR) module applies BEV RoI pooling, self-attention, and an auxiliary 3D point reconstruction task with Ground Truth Optimization (GTO) to refine proposals. The paper also proposes Multi-Agent Collaborative Data Augmentation (MCDA), combining local rotation, global scaling, and per-agent flipping. Experiments on DAIR-V2X, V2V4Real, and V2X-Real report state-of-the-art results, with headline gains of 10.86, 10.34, and 10.18 points in 3D AP@0.7.

Significance. The proposed architecture is well motivated and the module-level ablations (Tables 3–5) provide useful evidence for the individual components. If the reported comparisons are obtained under a controlled protocol, the work would be a substantial advance in broadcast-based collaborative 3D detection and would strengthen the case for explicitly handling 3D spatial misalignment during fusion. The paper provides a code link and clear architectural diagrams. However, the reliability of the headline SOTA claim currently depends on baseline comparability that the manuscript does not establish: the DPTP column in Table 6 appears to reproduce prior published numbers verbatim, and the DAIR-V2X evaluation uses CoAlign-completed annotations without clarifying whether baselines used the same labels. These issues must be resolved before the performance claims can be accepted.

major comments (4)
  1. [§4.2, Tables 1–2, 6] Section 4.2 says 'the models used for comparison are augmented with DPTP', but Table 6's DPTP column reports exactly the same 3D AP values as Table 1 and as the original papers for DI-V2X (72.54/39.24), ERMVP (68.61/37.51), DSRC (67.95/36.08), and CoSDH (67.95/36.78). Exact agreement to two decimals is implausible if these models were retrained under a common protocol. Combined with §4.1's use of CoAlign-completed DAIR-V2X annotations, this means the baseline margins in Tables 1–2 are not established as apples-to-apples. Please retrain all baselines under the identical data/range/augmentation protocol, or identify which numbers are taken from prior papers and explain why they remain comparable. Table 6's conclusion that MCDA 'consistently improves' prior methods is unsupported unless both its DPTP and MCDA columns are actually measured.
  2. [Abstract; §4.3; Table 1] The headline improvement on DAIR-V2X is internally inconsistent. The abstract states 10.86% 3D AP@0.7 improvement, which equals CoGoal3D minus CoAlign (50.55−39.69), while §4.3 says CoGoal3D outperforms DI-V2X by 11.31% (50.55−39.24). Please state the reference baseline and use one number throughout. Similarly, the abstract's 10.18% gain on V2X-Real corresponds to the gap over ERMVP (54.64−44.46), but §4.3 compares with DSRC and reports 11.97%; the comparison should be made against the best prior method in Table 2.
  3. [§3.4; Table 5] The MCDA local rotation is underspecified. The text says 'a unified local rotation around the ground truth objects' but does not define how points are associated with objects, what happens to background or overlapping points, or how the per-object rotation is applied when multiple objects overlap. Since Table 5 attributes an 8.67% 3D AP@0.7 improvement to MCDA, this step needs a precise algorithmic description or pseudocode to be reproducible.
  4. [§4.3–4.4] All reported AP numbers appear to be from a single run. Some ablation differences are small (e.g., removing GTO in Table 4 costs 0.98% 3D AP@0.7), and without repeated runs or variance estimates it is unclear whether such differences are meaningful. Reporting mean±std over at least three seeds for the main tables and ablations would materially strengthen the empirical claims.
minor comments (4)
  1. [Table 1] Table 1 lists FPS values, but no measurement protocol or hardware is specified; these may come from different papers and are not directly comparable. Please state the setup or remove the column.
  2. [Figure 7] Figure 7 shows robustness curves but lacks axis labels/units and a quantitative protocol (noise type, latency simulation, number of runs); consider moving this to a table.
  3. [§4.1; Reference [17]] There are typographical issues, e.g., 'V2X-Realtestsets' in §4.1 and 'alargs-scale' in reference [17]; please proofread.
  4. [§4.2] The paper does not report communication cost (bytes/latency) despite emphasizing the broadcast paradigm; adding this would support the efficiency claim.

Circularity Check

0 steps flagged

No circular derivation: CoGoal3D is an empirical supervised detection pipeline whose losses and ground truths are external, not self-referential.

full rationale

The paper proposes a two-stage collaborative 3D detector trained with standard supervised losses: RPN loss, refinement loss, and Chamfer-distance reconstruction loss (Eq. 7). The reconstruction supervision is generated from annotated 3D bounding boxes and raw LiDAR points via Ground Truth Optimization, not from the model's own predictions. There is no equation in which an output is defined in terms of the quantity it is claimed to predict, and no fitted parameter is renamed as a prediction. The claimed contribution is architectural (3D-aware fusion and refinement) plus a data-augmentation strategy, evaluated on public real-world benchmarks. The ablation studies compare the model with and without its own components, which is a standard empirical attribution, not a circular derivation. Concerns about baseline comparability in Tables 1, 2, and 6 are validity/reproducibility issues, not circularity: they do not show that the derived result is equivalent to its inputs by construction. The paper does not invoke self-citations as load-bearing mathematical premises or uniqueness theorems. Therefore no circular step is identified, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

This is an empirical deep-learning paper; there is no closed-form physical derivation and no fitted scientific constant. The learned network weights are optimized on training folds and are not ledger entries. The entries above are the hand-set hyperparameters and domain assumptions that the reported performance depends on, none of which are machine-checked.

free parameters (3)
  • 3D-AGF multiscale configuration (3 layers, 8 heads, 9 sampling points)
    Chosen in Section 4.2 without sensitivity analysis; central fusion module capacity depends on these values.
  • RoI grid size G=6 and reconstructed points per pillar Np=6
    Section 4.2; no ablation of these values is reported.
  • MCDA augmentation ranges (flip 50%, rotation ±π/20, scale 0.95–1.05)
    Section 4.2; hand-selected, not swept, and the MCDA gains could be sensitive to these ranges.
axioms (5)
  • domain assumption 3D position encoding plus deformable cross-attention can correct height/attitude misalignment among agents
    Core premise of Section 3.2; no theoretical guarantee, supported only indirectly by ablations in Tables 3 and 4.
  • domain assumption GTO-projected collaborative points are a valid 3D supervision signal for the reconstruction head
    Section 3.3; assumes calibration and annotation matching are accurate enough that projecting agent points into ego GT boxes improves geometry learning.
  • domain assumption MCDA augmentation does not shift the evaluation distribution and reduces information loss as claimed
    Section 3.4; supported by Figure 6 and downstream AP, but no quantitative point-level information-loss measurement is given.
  • domain assumption PointPillars BEV features preserve sufficient 3D structure for auxiliary point reconstruction to help box refinement
    Design premise of the two-stage pipeline in Sections 3.1 and 3.3; no formal justification beyond the ablation results.
  • standard math Chamfer distance and Hungarian matching behave as published and are appropriate for the proposed modules
    Used in Sections 3.3 and 3.5 as standard differentiable losses and assignment algorithms; not re-derived.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CoGoal3D: Collaborative 3D Object Detection with 3D-Aware Fusion and Refinement." pith.science (2026). https://pith.science/paper/PC3L3PML

@misc{pith2026260719036,
  author       = {Pith},
  title        = {Pith review of: CoGoal3D: Collaborative 3D Object Detection with 3D-Aware Fusion and Refinement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PC3L3PML}},
  note         = {Machine review of arXiv:2607.19036}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

V2X collaborative object detection features overcoming the limitations of single-vehicle systems by aggregating environmental features from multiple collaborative agents. However, existing mainstream V2X perception methods mainly focus on 2D BEV object detection. When 3D detection task is concerned, inferior results are obtained because they ignore the 3D spatial misalignment caused by differing height and attitude among the collaborators. In this paper, we propose a novel collaborative 3D object detection framework called CoGoal3D, which extracts and refines the 3D feature gradually in a two-stage pipeline. In the first stage, a multiscale 3D-aware global fusion module is designed to mitigate the 3D spatial misalignment. The resulting proposals are then refined in the second stage with an auxiliary task of 3D point reconstruction. An effective multi-agent collaborative data augmentation strategy is further proposed to enrich the training data while minimizing information loss. Extensive experiments on public real-world datasets demonstrate that our CoGoal3D achieves new state-of-the-art performance, with 3D AP@0.7 improvements of 10.86%, 10.34%, and 10.18% on the DAIR-V2X, V2V4Real, and V2X-Real datasets, respectively. Code is available at https://github.com/Megalo-f/CoGoal3D.

Figures

Figures reproduced from arXiv: 2607.19036 by Dongping Zhang, Eryun Liu, Kai Wang, Peng Xu, Tianyu Pu, Yong Ding, Zhihao Yang, Zhiyu Xiang.

Figure 1
Figure 1. Figure 1: Difference between the mainstream methods and ours. (a) Mainstream broadcast-based methods perform feature fusion only on 2D BEV space. (b) Ours is a two-stage pipeline, enhancing the 3D alignment by 3D-aware global fusion and reconstruction-guided local refinement respectively. a highly attractive solution to address these limitations by leveraging multi-view information through V2X. Currently most of the… view at source ↗
Figure 2
Figure 2. Figure 2: The overall architecture of the proposed CoGoal3D. The dashed boxes and lines indicate the components used exclusively during training. Further details of these components are illustrated in Section 3. Collaborative Data Augmentation (MCDA) is applied to the point clouds and poses of all collaborative agents. In the stage 1, the input LiDAR points are passed through shared 3D back￾bones, from which each ag… view at source ↗
Figure 3
Figure 3. Figure 3: The architecture of the pro￾posed 3D-Aware Global Fusion (3D￾AGF) module. It explicitly incorpo￾rates 3D spatial information to align collaborative features via 3D position encoding and 3D-aware deformable cross attention. (a) (b) (c) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the proposed Multi-Agent Collaborative Data Aug￾mentation (MCDA) strategy. It applies a specific sequence of local and global augmentations, maximizing data diver￾sity while minimizing information loss. 𝑥 𝑦 𝑥 𝑦 𝜃 𝜃 (b) MCDA (a) DPTP Data Rotation Flipping 𝑥 𝑦 𝑥 𝑦 Ego Ego 𝑥 𝑦 𝑥 𝑦 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Robustness evaluation against localization error (left) and transmission latency (right) on the DAIR-V2X val set. Robust Evaluation. We further evaluate the robustness of CoGoal3D against localization error and transmission latency. As shown in [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative results on DAIR-V2X val set. The first row shows the 3D zoom-in views of the blue windows in the second row of BEV views. Green and red bounding boxes denote the 3D object ground truths and the detection results, respectively. 4.5 Qualitative Evaluation A representative qualitative comparison between our CoGoal3D and other meth￾ods on the DAIR-V2X dataset is shown in [PITH_FULL_IMAGE:figures/f… 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

28 extracted references · 4 linked inside Pith

  1. [1]

    In: Proceedings of the AAAI conference on artificial intelligence

    Deng, J., Shi, S., Li, P., Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high per- formance voxel-based 3d object detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 1201–1209 (2021)

    2021

  2. [2]

    Advances in neural information processing systems35, 4874–4886 (2022)

    Hu, Y., Fang, S., Lei, Z., Zhong, Y., Chen, S.: Where2comm: Communication- efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems35, 4874–4886 (2022)

    2022

  3. [3]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  4. [4]

    Naval research logistics quarterly2(1-2), 83–97 (1955)

    Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly2(1-2), 83–97 (1955)

    1955

  5. [5]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12697–12705 (2019)

    2019

  6. [6]

    In:ProceedingsoftheAAAIConferenceonArtificialIntelligence.vol.38,pp.3208– 3215 (2024)

    Li, X., Yin, J., Li, W., Xu, C., Yang, R., Shen, J.: Di-v2x: Learning domain- invariantrepresentationforvehicle-infrastructurecollaborative3dobjectdetection. In:ProceedingsoftheAAAIConferenceonArtificialIntelligence.vol.38,pp.3208– 3215 (2024)

    2024

  7. [7]

    Advances in Neural Information Processing Systems34, 29541–29552 (2021)

    Li, Y., Ren, S., Wu, P., Chen, S., Feng, C., Zhang, W.: Learning distilled collabora- tion graph for multi-agent perception. Advances in Neural Information Processing Systems34, 29541–29552 (2021)

    2021

  8. [8]

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object detection.In:ProceedingsoftheIEEEinternationalconferenceoncomputervision. pp. 2980–2988 (2017)

    2017

  9. [9]

    In: 2023 IEEE International Conference on Robotics and Automation (ICRA)

    Lu, Y., Li, Q., Liu, B., Dianati, M., Feng, C., Chen, S., Wang, Y.: Robust collab- orative 3d object detection in presence of pose errors. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 4812–4818. IEEE (2023)

    2023

  10. [10]

    Advances in neural information processing sys- tems32(2019) 16 Z

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. Advances in neural information processing sys- tems32(2019) 16 Z. Yang et al

    2019

  11. [11]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652–660 (2017)

    2017

  12. [12]

    Advances in neural information processing systems30(2017)

    Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learn- ing on point sets in a metric space. Advances in neural information processing systems30(2017)

    2017

  13. [13]

    arXiv preprint arXiv:2302.13301 (2023)

    Shi, G., Li, R., Ma, C.: Pillar r-cnn for point cloud 3d object detection. arXiv preprint arXiv:2302.13301 (2023)

    Pith/arXiv arXiv 2023

  14. [14]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10529–10538 (2020)

    2020

  15. [15]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Shi, S., Wang, X., Li, H.: Pointrcnn: 3d object proposal generation and detection from point cloud. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 770–779 (2019)

    2019

  16. [16]

    Wang,T.H.,Manivasagam,S.,Liang,M.,Yang,B.,Zeng,W.,Urtasun,R.:V2vnet: Vehicle-to-vehiclecommunicationforjointperceptionandprediction.In:Computer Vision–ECCV 2020: 16th European Conference,Glasgow,UK, August23–28, 2020, Proceedings, Part II 16. pp. 605–621. Springer (2020)

    2020

  17. [17]

    In: European Conference on Computer Vision

    Xiang, H., Zheng, Z., Xia, X., Xu, R., Gao, L., Zhou, Z., Han, X., Ji, X., Li, M., Meng,Z.,etal.:V2x-real:alargs-scaledatasetforvehicle-to-everythingcooperative perception. In: European Conference on Computer Vision. pp. 455–470. Springer (2024)

    2024

  18. [18]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Xu, J., Zhang, Y., Cai, Z., Huang, D.: Cosdh: Communication-efficient collabora- tive perception via supply-demand awareness and intermediate-late hybridization. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 6834–6843 (2025)

    2025

  19. [19]

    IEEE transactions on pattern analysis and machine intelligence (2024)

    Xu, R., Chen, C.J., Tu, Z., Yang, M.H.: V2x-vitv2: Improved vision transform- ers for vehicle-to-everything cooperative perception. IEEE transactions on pattern analysis and machine intelligence (2024)

    2024

  20. [20]

    arXiv preprint arXiv:2207.02202 (2022)

    Xu, R., Tu, Z., Xiang, H., Shao, W., Zhou, B., Ma, J.: Cobevt: Cooperative bird’s eye view semantic segmentation with sparse transformers. arXiv preprint arXiv:2207.02202 (2022)

    Pith/arXiv arXiv 2022

  21. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Xu, R., Xia, X., Li, J., Li, H., Zhang, S., Tu, Z., Meng, Z., Xiang, H., Dong, X., Song, R., et al.: V2v4real: A real-world large-scale dataset for vehicle-to-vehicle co- operative perception. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13712–13722 (2023)

    2023

  22. [22]

    In: European confer- ence on computer vision

    Xu, R., Xiang, H., Tu, Z., Xia, X., Yang, M.H., Ma, J.: V2x-vit: Vehicle-to- everything cooperative perception with vision transformer. In: European confer- ence on computer vision. pp. 107–124. Springer (2022)

    2022

  23. [23]

    Sen- sors18(10), 3337 (2018)

    Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sen- sors18(10), 3337 (2018)

    2018

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yu, H., Luo, Y., Shu, M., Huo, Y., Yang, Z., Shi, Y., Guo, Z., Li, H., Hu, X., Yuan, J., et al.: Dair-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21361–21370 (2022)

    2022

  25. [25]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Zhang, J., Wang, Y., Qian, L., Sun, P., Li, Z., Jiang, S., Liu, M., Song, L.: Dsrc: Learning density-insensitive and semantic-aware collaborative representa- tion against corruptions. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 9942–9950 (2025) CoGoal3D: Collaborative 3D Object Detection with 3D-Aware Fusion 17

    2025

  26. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhang, J., Yang, K., Wang, Y., Wang, H., Sun, P., Song, L.: Ermvp: Communication-efficient and collaboration-robust multi-vehicle perception in chal- lenging environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12575–12584 (2024)

    2024

  27. [27]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4490–4499 (2018)

    2018

  28. [28]

    arXiv preprint arXiv:2010.04159 (2020)

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

    Pith/arXiv arXiv 2010

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.