Pith. sign in

REVIEW 5 major objections 7 minor 51 references

Non-Overlap-Aware Egocentric Pose Estimation for Collaborative Perception in Connected Autonomy

T0 review · 5 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read A two-stage graph-matching and cross-attention approach lets connected vehicles detect non-overlapping views, estimate relative pose from a 27 KB graph packet, and beat the leading baseline by 53.9% in position and 78.6% in rotation.

desk verdict Solid engineering, a new task framing, and a real bandwidth win, but the overlap-decision failure mode and thin statistics keep the SOTA claim conditional. read the letter →

arxiv 2506.14180 v2 pith:Q4BPC2KW submitted 2025-06-17 cs.RO

classification cs.RO
keywords egocentricposeestimationcollaborativeperceptiongraphmatchingnon-overlapdetectionconnectedautonomousvehiclescommunicationefficiencyposition-awarecross-attentionmulti-robotsystems
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

Connected and autonomous vehicles need to know the relative pose of a teammate before merging perceptions, but their cameras often see completely different scenes containing similar objects, and sharing raw images exceeds V2X bandwidth. This paper proposes NOPE, a hierarchical learning method that first decides whether two views overlap by matching graphs of detected objects, and only then estimates the teammate's position and rotation with a position-aware cross-attention network. The central claim is that this two-level design both avoids wrong pose estimates on non-overlapping views and cuts shared data to about 27 KB per exchange. In simulation, NOPE reports 53.9 percent lower position error and 78.6 percent lower rotation error than the strongest baseline, and it detects non-overlapping views correctly in about 80 percent of simulated cases.

What carries the argument

The central object is an attributed graph per observation: nodes are detected objects with LVM-extracted visual features, edges are formed by Delaunay triangulation, and edge attributes encode Euclidean distance between objects. The argument is carried by two learned stages. High-level graph matching treats non-overlap detection as a correspondence problem, using a graph consensus theorem to refine a similarity matrix and the Hungarian algorithm to enforce one-to-one matching. Low-level pose estimation uses a position-aware cross-attention graph network that concatenates node embeddings with learnable position embeddings, applies cross-attention between the two graphs, and pools the result with attention gates to regress position and rotation. Together these make the shared representation a graph packet of about 27 KB, compared with 0.75 MB for the CoViS-Net baseline and 2.3 MB for SuperGlue.

What would settle it

Take a pair of genuinely overlapping views in which a shared object is missing from YOLOv5's detections (for example, a small or heavily occluded vehicle), run NOPE, and check whether it declares the views non-overlapping and skips pose estimation; if it does, the reported non-overlap accuracy is dominated by detector recall rather than by true visual overlap.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that relative pose estimation between robots can be made overlap-aware and bandwidth-aware by compressing each robot's observation into an object graph rather than sharing images or point clouds. The high-level stage solves a graph matching problem over nodes with features from large visual models and edge weights from Delaunay-triangulated 3D object positions; if no correspondence survives thresholding, the views are declared non-overlapping and pose estimation is skipped. The low-level stage then pools graph embeddings through cross-attention and an attention gate to regress a position and a quaternion rotation. NOPE is evaluated in high-fidelity simulation (30,277 data instances) and on a real-world multi-vehicle dataset (201 data instances), and it outperforms the six compared methods in pose error while transmitting about 1/96 of the data size of the best baseline.

Load-bearing premise

The method assumes the object detector and depth estimator reliably find every shared object, because two views that really overlap but contain no detected shared object are classified as non-overlapping and no pose is produced.

Editorial extensions

If this is right

  • If the simulation results transfer, connected vehicles can avoid wrong relative poses that arise when two cameras see similar but non-overlapping scenes, because the pose branch only runs after a correspondence-based overlap check.
  • The same graph packet can be exchanged under realistic V2X bandwidth: 27 KB per pair is roughly 96 times smaller than the CoViS-Net baseline's 0.75 MB.
  • Because NOPE does not need GPS or a shared map at inference time, it applies to GPS-denied urban areas where GNSS-based collaborative perception fails.
  • On the real-world dataset, NOPE keeps the lowest pose errors among all compared methods without any fine-tuning, which supports cross-domain generalizability.

Reading between the lines

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

  • The paper leaves untouched a safety consequence of its binary overlap rule: two genuinely overlapping views with no detected shared object are silently classified as non-overlapping and no pose is produced, so a confidence score on the 'no correspondence' verdict would make the system far safer in dense scenes.
  • Because the shared artifact is a small graph rather than raw sensor data, the same correspondence matrix could in principle be reused for tasks beyond pose, such as cooperative object association or lightweight map merging, without re-detecting overlap.
  • The reported 80.39 percent non-overlap detection accuracy leaves roughly one case in five mislabeled, so tracking correspondences over successive frames or fusing multiple views could materially improve the overlap decision before the pose branch runs.
  • The edge attributes inherit all the noise of the depth estimator, so an uncertainty-aware edge encoding would be a natural testable extension of the graph representation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 7 minor

Summary. The paper proposes NOPE, a hierarchical learning framework for non-overlap-aware egocentric pose estimation between pairs of connected vehicles. Each vehicle's observation is represented as a graph whose nodes are detected objects with visual features and whose edges are Delaunay triangulations with depth-derived distances. A high-level deep graph matching network (CoID) identifies correspondences between two views; if no correspondence survives thresholding, the views are declared non-overlapping and pose estimation is skipped. Otherwise, a low-level position-aware cross-attention network estimates the relative position and rotation from graph-level embeddings. The method is evaluated in CARLA/SUMO simulation (30,277 instances) and on the MARS real-world dataset (201 instances), reporting that NOPE outperforms CoViS-Net by 53.9% in position error and 78.6% in rotation error in simulation, with 96.48% reduction in shared data size, and achieves 80.39% non-overlap detection accuracy in simulation.

Significance. If the reported results hold, NOPE is a meaningful contribution to collaborative perception for connected autonomy: it is the first learning-based egocentric pose estimation method that explicitly handles non-overlapping views while respecting realistic V2X bandwidth constraints. The graph-based representation and hierarchical design are sensible, and the claimed improvements over strong baselines are substantial. The paper also honestly acknowledges its limitations, including inability to estimate poses for completely non-overlapping views and lack of global consistency for teams larger than two. However, the central claims rest on a small real-world evaluation, a potentially test-set-selected threshold, and an unmeasured failure mode in which detection failures cause overlapping views to be classified as non-overlapping. These issues need to be addressed before the claims can be fully trusted.

major comments (5)
  1. [Section III-B and Table II] The non-overlap decision is made solely by checking whether the Hungarian-assignment matrix Y has any nonzero entries: if sum Y = 0, the views are declared non-overlapping and pose estimation is skipped. This makes the claimed non-overlap-awareness conditional on the object detector and depth estimator producing nodes for all covisible objects. The paper does not measure the failure mode where genuinely overlapping views have no detected shared object due to occlusion, low resolution, or detector failure. Table II reports only NDA (80.39%) on the simulation test set and does not provide a confusion matrix or the false non-overlap rate on truly overlapping pairs, even though ground-truth correspondences are available in simulation. Please report the fraction of overlapping pairs that are misclassified as non-overlapping in the simulation test set and discuss the impact of this failure mode on the claimed new capability.
  2. [Section IV-D and Figure 5] The threshold tau in Eq. (7) is selected based on the analysis shown in Figure 5, where the highest NDA is reported for tau in [0.6, 0.7]. The paper does not state whether this sweep was performed on a training/validation split or on the test set. Since the test set is used to report the final NDA in Table II, threshold selection on the test set would make the result optimistic and potentially not reproducible. Please specify the data split used for this analysis and, if necessary, re-report the results with tau fixed using only training/validation data.
  3. [Tables I and II] All quantitative results are single point estimates without error bars, standard deviations, or significance tests. This is concerning because the training and inference pipeline includes stochastic elements (dropout, random matrix J in Eq. (6), and stochastic optimization) and because the claimed improvements are large (e.g., 53.87% position improvement over CoViS-Net). Please provide means and standard deviations over multiple runs or seeds, and state whether the reported improvements are statistically significant.
  4. [Section IV-A and Eq. (9)] The high-level loss in Eq. (9) is defined using S_hat, which is obtained through the indicator function epsilon in Eq. (7). This indicator function is non-differentiable, so it is unclear how the CoID network is trained end-to-end with this loss. The paper does not describe any continuous relaxation or surrogate loss used during training. Please clarify the training procedure, as this directly affects the validity of the CoID and non-overlap detection results reported in Table II.
  5. [Section IV-C and Table I] The real-world evaluation uses only 201 data instances, with no ground-truth correspondences and no fine-tuning of the model trained in simulation. The real-world position error improvement over CoViS-Net is only 6.32% and the rotation error improvement is 28.65%, both computed on a single, small test set. The paper states that NOPE 'continuously maintains a low pose error' but does not provide confidence intervals, per-instance breakdowns, or an analysis of sensitivity to the specific test instances. Please provide additional statistical support (e.g., bootstrap confidence intervals) and discuss the variability of the results.
minor comments (7)
  1. [Abstract and Introduction] There are minor typos: 'satifying' should be 'satisfying' in the abstract, and 'an unified' should be 'a unified' in the first paragraph of the introduction.
  2. [Eq. (6)] The definition of the graph difference matrix D appears to have a dimension inconsistency: S^T Ψ(J, A) and Ψ(S^T J, A') are not obviously the same shape, and the statement 'Di,j = 0' when the graphs are the same is presented without explaining how the index (i,j) is defined. Please revise the notation and clarify the dimensions and the graph-consensus argument.
  3. [Section III-B, Eq. (7) and Eq. (8)] The indicator function epsilon binarizes the similarity matrix before the Hungarian algorithm, which is an unusual formulation because the Hungarian algorithm is typically applied to a continuous score matrix. Please clarify whether the binarization is applied only at inference and whether a continuous version is used for training the high-level network.
  4. [Section IV-A] The experimental setup says 'we utilize both CARLA and SUMO to create five connected autonomous driving (CAD) scenarios,' but the five scenarios are never listed or described. Please provide a brief description of the scenarios and how they differ.
  5. [Figure 5] The analysis of tau is difficult to read: the axes are not labeled, the metric names are abbreviated, and the figure does not show any confidence intervals. Please improve the figure and include axis labels and units.
  6. [Table I and Table II] The baseline NOPEhigh is presented in Table II as an ablation of the high-level component, but it is not included in Table I. Including NOPEhigh in Table I would help quantify the contribution of the low-level pose estimation network.
  7. [Section IV-A, metrics] The definition of NDA states it is 'the ratio of correctly detected non-overlapping observation pairs to the total number of observation pairs,' which reads as if only non-overlapping pairs are counted. Please clarify whether NDA is a binary classification accuracy over both overlapping and non-overlapping pairs, or something else.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: predictions come from learned models evaluated on held-out and sim-to-real data; self-citations are baselines, not load-bearing premises.

full rationale

I examined the claimed derivation chain. The high-level CoID network is trained with ground-truth correspondences from CARLA via Eq. (9), and the low-level pose network is trained with ground-truth poses via Eq. (17); both are evaluated on held-out simulation splits and on MARS without fine-tuning, so no fitted parameter is renamed as a prediction. Non-overlap detection is explicitly defined as the absence of correspondences: Section III-B states 'if the sum of all elements in Σ_{i,j} Y = 0, it indicates that there is no overlap between two robots’ views due to the lack of correspondences.' That is a design choice for the system's behavior, not a derivation that reduces a predicted quantity to its own input. The graph-consensus difference matrix in Eq. (6) is justified by an external theorem [40] (Fey et al., DGMC), not by a self-citation of the authors. The authors' prior graph-matching works appear as experimental baselines (BDGM, DMGM) and related work, not as premises of the NOPE derivation. The known limitations—dependence on object detection/depth quality and abstention for fully non-overlapping views—are robustness concerns, not circularities. I found no equation, threshold, or baseline that is equivalent by construction to the claimed outputs, so the paper is self-contained against the identified circularity patterns.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the graph abstraction (detected objects plus Delaunay edges), the learned correspondences for overlap detection, and the supervised pose predictor. The only explicit hand-chosen numeric constant appears to be the detection threshold tau; the architecture hyperparameters are standard manual choices. The method assumes detector/depth reliability and that overlap can be reduced to existence of correspondences.

free parameters (2)
  • Threshold tau for correspondence binarization = reported best range [0.6, 0.7]
    Eq. (7) uses tau to convert similarity plus graph difference into binary matches; Section IV-D selects tau by inspecting NDA and F1 curves without clear validation/test separation.
  • Architecture and training hyperparameters = L=2 (high) and L=4 (low), heads=4, d=256, dropout 0.5/0.2, lr=0.001, 150 epochs
    Section IV-A sets these by hand; they affect the reported performance and are not derived from theory.
assumptions (6)
  • standard math Hungarian algorithm solves the assignment problem in Eq. (8) optimally
    Used to enforce one-to-one correspondences; standard combinatorial optimization.
  • standard math Graph consensus theorem (cited as [40]) makes D vanish for identical graphs
    Eq. (6) relies on this theorem to justify the graph difference update.
  • domain assumption YOLOv5 and DepthAnythingV2 provide accurate object detection and depth for graph construction
    Section IV-A builds nodes from YOLOv5 detections and edge attributes from DepthAnythingV2; failure here corrupts the graph and the non-overlap decision.
  • domain assumption Overlap is fully determined by the existence of at least one learned correspondence (sum Y >= 1)
    Section III-B declares non-overlap when no correspondences survive; this equates geometric overlap with matching success.
  • domain assumption CARLA and GPS ground truth poses and correspondences are error-free for supervision
    Training and evaluation in Section IV-A rely on these labels; label noise is not modeled.
  • domain assumption The two-robot, front-camera setup with a 7.2 Mbps V2X cap is representative of connected autonomy
    Motivation in Section I and baseline bandwidth comparison in Table I assume this operating regime.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Non-Overlap-Aware Egocentric Pose Estimation for Collaborative Perception in Connected Autonomy." pith.science (2026). https://pith.science/paper/Q4BPC2KW

@misc{pith2026250614180,
  author       = {Pith},
  title        = {Pith review of: Non-Overlap-Aware Egocentric Pose Estimation for Collaborative Perception in Connected Autonomy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q4BPC2KW}},
  note         = {Machine review of arXiv:2506.14180}
}
read the original abstract

Egocentric pose estimation is a fundamental capability for multi-robot collaborative perception in connected autonomy, such as connected autonomous vehicles. During multi-robot operations, a robot needs to know the relative pose between itself and its teammates with respect to its own coordinates. However, different robots usually observe completely different views that contains similar objects, which leads to wrong pose estimation. In addition, it is unrealistic to allow robots to share their raw observations to detect overlap due to the limited communication bandwidth constraint. In this paper, we introduce a novel method for Non-Overlap-Aware Egocentric Pose Estimation (NOPE), which performs egocentric pose estimation in a multi-robot team while identifying the non-overlap views and satifying the communication bandwidth constraint. NOPE is built upon an unified hierarchical learning framework that integrates two levels of robot learning: (1) high-level deep graph matching for correspondence identification, which allows to identify if two views are overlapping or not, (2) low-level position-aware cross-attention graph learning for egocentric pose estimation. To evaluate NOPE, we conduct extensive experiments in both high-fidelity simulation and real-world scenarios. Experimental results have demonstrated that NOPE enables the novel capability for non-overlapping-aware egocentric pose estimation and achieves state-of-art performance compared with the existing methods. Our project page at https://hongh0.github.io/NOPE/.

Figures

Figures reproduced from arXiv: 2506.14180 by the authors.

Figure 1
Figure 1. A motivating scenario for egocentric pose estimation in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our NOPE framework. NOPE represents the observation of each robot as a graph. The high-level NOPE [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results on CoID and egocentric pose estimation from both simulation and real-world scenarios. The first [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparisons of CoID for non-overlap detection. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 40 canonical work pages

  1. [1]

    Do we run large-scale multi-robot systems on the edge? more evidence for two-phase performance in system size scaling,

    J. Kuckling, R. Luckey, V . Avrutin, A. Vardy, A. Reina, and H. Hamann, “Do we run large-scale multi-robot systems on the edge? more evidence for two-phase performance in system size scaling,” in ICRA, 2024

  2. [2]

    Relative localizability and localization for multi-robot systems,

    L. Chen, C. Liang, S. Yuan, M. Cao, and L. Xie, “Relative localizability and localization for multi-robot systems,” TRO, pp. 1–19, 2025

  3. [3]

    Fault-tolerant rendezvous of multirobot systems,

    H. Park and S. A. Hutchinson, “Fault-tolerant rendezvous of multirobot systems,” TRO, vol. 33, no. 3, pp. 565–582, 2017

  4. [4]

    Towards vehicle-to-everything autonomous driving: A survey on collaborative perception,

    S. Liu, C. Gao, Y . Chen, X. Peng, X. Kong, K. Wang, R. Xu, W. Jiang, H. Xiang, J. Ma, et al. , “Towards vehicle-to-everything autonomous driving: A survey on collaborative perception,” arXiv preprint arXiv:2308.16714, 2023

  5. [5]

    Collaborative perception for connected and autonomous driving: Challenges, possible solutions and opportunities,

    S. Hu, Z. Fang, Y . Deng, X. Chen, and Y . Fang, “Collaborative perception for connected and autonomous driving: Challenges, possible solutions and opportunities,” arXiv preprint arXiv:2401.01544 , 2024

  6. [6]

    Edge-assisted multi-robot visual-inertial slam with efficient communication,

    X. Liu, S. Wen, J. Zhao, T. Z. Qiu, and H. Zhang, “Edge-assisted multi-robot visual-inertial slam with efficient communication,” IEEE Transactions on Automation Science and Engineering , 2024

  7. [7]

    A survey on active simultaneous localization and mapping: State of the art and new frontiers,

    J. A. Placed, J. Strader, H. Carrillo, N. Atanasov, V . Indelman, L. Carlone, and J. A. Castellanos, “A survey on active simultaneous localization and mapping: State of the art and new frontiers,” TRO, vol. 39, no. 3, pp. 1686–1705, 2023

  8. [8]

    S3e: A multi-robot multimodal dataset for collaborative slam,

    D. Feng, Y . Qi, S. Zhong, Z. Chen, Q. Chen, H. Chen, J. Wu, and J. Ma, “S3e: A multi-robot multimodal dataset for collaborative slam,” RAL, 2024

Show all 51 references
  1. [9]

    Lamp 2.0: A robust multi-robot slam system for operation in challenging large-scale underground environments,

    Y . Chang, K. Ebadi, C. E. Denniston, M. F. Ginting, A. Rosinol, A. Reinke, M. Palieri, J. Shi, A. Chatterjee, B. Morrell, et al., “Lamp 2.0: A robust multi-robot slam system for operation in challenging large-scale underground environments,” RAL, vol. 7, no. 4, pp. 9175– 9182, 2022

  2. [10]

    From simulations to reality: enhancing multi-robot explo- ration for urban search and rescue,

    G. S. Kashyap, D. Mahajan, O. C. Phukan, A. Kumar, A. E. Brownlee, and J. Gao, “From simulations to reality: enhancing multi-robot explo- ration for urban search and rescue,” arXiv preprint arXiv:2311.16958 , 2023

  3. [11]

    Any way you look at it: Semantic crossview localization and mapping with lidar,

    I. D. Miller, A. Cowley, R. Konkimalla, S. S. Shivakumar, T. Nguyen, T. Smith, C. J. Taylor, and V . Kumar, “Any way you look at it: Semantic crossview localization and mapping with lidar,” RAL, vol. 6, no. 2, pp. 2397–2404, 2021

  4. [12]

    Fusionvlad: A multi-view deep fusion networks for viewpoint-free 3d place recognition,

    P. Yin, L. Xu, J. Zhang, and H. Choset, “Fusionvlad: A multi-view deep fusion networks for viewpoint-free 3d place recognition,” RAL, vol. 6, no. 2, pp. 2304–2310, 2021

  5. [13]

    Loftr: Detector-free local feature matching with transformers,

    J. Sun, Z. Shen, Y . Wang, H. Bao, and X. Zhou, “Loftr: Detector-free local feature matching with transformers,” in CVPR, 2021

  6. [14]

    Deepmatcher: a deep transformer-based network for robust and accurate local feature matching,

    T. Xie, K. Dai, K. Wang, R. Li, and L. Zhao, “Deepmatcher: a deep transformer-based network for robust and accurate local feature matching,” Expert Systems with Applications , 2024

  7. [15]

    Se (3) diffusion model-based point cloud registration for robust 6d object pose estimation,

    H. Jiang, M. Salzmann, Z. Dang, J. Xie, and J. Yang, “Se (3) diffusion model-based point cloud registration for robust 6d object pose estimation,” NeurIPS, 2024

  8. [16]

    Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment,

    J. Wang, C. Rupprecht, and D. Novotny, “Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment,” in CVPR, 2023

  9. [17]

    A lte-direct broadcast mechanism for periodic vehicular safety communications,

    L. Gallo and J. H ¨arri, “A lte-direct broadcast mechanism for periodic vehicular safety communications,” in IEEE VNC, 2013

  10. [18]

    Asynchronous collaborative localization by integrating spatiotemporal graph learning with model-based estimation,

    P. Gao, B. Reily, R. Guo, H. Lu, Q. Zhu, and H. Zhang, “Asynchronous collaborative localization by integrating spatiotemporal graph learning with model-based estimation,” in ICRA, 2022

  11. [19]

    Surfacenet: An end-to- end 3d neural network for multiview stereopsis,

    M. Ji, J. Gall, H. Zheng, Y . Liu, and L. Fang, “Surfacenet: An end-to- end 3d neural network for multiview stereopsis,” in ICCV, 2017

  12. [20]

    Multi-view sensor fusion by integrating model-based estimation and graph learning for collaborative object localization,

    P. Gao, R. Guo, H. Lu, and H. Zhang, “Multi-view sensor fusion by integrating model-based estimation and graph learning for collaborative object localization,” in ICRA, 2021

  13. [21]

    Learning interaction-aware trajectory predictions for decentralized multi-robot motion planning in dynamic environments,

    H. Zhu, F. M. Claramunt, B. Brito, and J. Alonso-Mora, “Learning interaction-aware trajectory predictions for decentralized multi-robot motion planning in dynamic environments,” RAL, vol. 6, no. 2, pp. 2256–2263, 2021

  14. [22]

    Who2com: Collaborative perception via learnable handshake commu- nication,

    Y .-C. Liu, J. Tian, C.-Y . Ma, N. Glaser, C.-W. Kuo, and Z. Kira, “Who2com: Collaborative perception via learnable handshake commu- nication,” in ICRA, 2020

  15. [23]

    When2com: Multi-agent perception via communication graph grouping,

    Y .-C. Liu, J. Tian, N. Glaser, and Z. Kira, “When2com: Multi-agent perception via communication graph grouping,” in CVPR, 2020

  16. [24]

    Multi-robot target detection and tracking: Taxonomy and survey,

    C. Robin and S. Lacroix, “Multi-robot target detection and tracking: Taxonomy and survey,” AuRo, 2016

  17. [25]

    Cooperative perception for 3d object detection in driving scenarios using infrastruc- ture sensors,

    E. Arnold, M. Dianati, R. de Temple, and S. Fallah, “Cooperative perception for 3d object detection in driving scenarios using infrastruc- ture sensors,” Intelligence Transportation System , vol. 23, no. 3, pp. 1852–1864, 2020

  18. [26]

    Where2comm: Communication-efficient collaborative perception via spatial confidence maps,

    Y . Hu, S. Fang, Z. Lei, Y . Zhong, and S. Chen, “Where2comm: Communication-efficient collaborative perception via spatial confidence maps,” NeurIPS, 2022

  19. [27]

    Object detection with discriminatively trained part-based models,

    D. Forsyth, “Object detection with discriminatively trained part-based models,” Computer, vol. 47, no. 02, pp. 6–7, 2014

  20. [28]

    A cooperative perception system robust to localization errors,

    Z. Song, F. Wen, H. Zhang, and J. Li, “A cooperative perception system robust to localization errors,” in IEEE IV, 2023, pp. 1–6

  21. [29]

    Array ppp-rtk: A high precision pose estimation method for outdoor scenarios,

    X. An, A. Bell ´es, F. G. Rizzi, L. H ¨osch, C. Lass, and D. Medina, “Array ppp-rtk: A high precision pose estimation method for outdoor scenarios,” IEEE ITS, vol. 25, no. 6, pp. 6223–6237, 2023

  22. [30]

    Deep masked graph matching for correspondence identification in collaborative perception,

    P. Gao, Q. Zhu, H. Lu, C. Gan, and H. Zhang, “Deep masked graph matching for correspondence identification in collaborative perception,” in ICRA, 2023

  23. [31]

    Uncertainty-aware correspondence identification for collaborative perception,

    P. Gao, Q. Zhu, and H. Zhang, “Uncertainty-aware correspondence identification for collaborative perception,” AuRo, vol. 47, no. 5, pp. 635–648, 2023

  24. [32]

    Visual, spatial, geometric-preserved place recognition for cross-view and cross-modal collaborative perception,

    P. Gao, J. Liang, Y . Shen, S. Son, and M. C. Lin, “Visual, spatial, geometric-preserved place recognition for cross-view and cross-modal collaborative perception,” in IROS. IEEE, 2023, pp. 11 079–11 086

  25. [33]

    Long-term loop closure detection through visual-spatial information preserving multi-order graph matching,

    P. Gao and H. Zhang, “Long-term loop closure detection through visual-spatial information preserving multi-order graph matching,” in AAAI, 2020

  26. [34]

    Superglue: Learning feature matching with graph neural networks,

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich, “Superglue: Learning feature matching with graph neural networks,” in CVPR, 2020

  27. [35]

    Diffglue: Diffusion-aided image feature match- ing,

    S. Zhang and J. Ma, “Diffglue: Diffusion-aided image feature match- ing,” in Proceedings of the 32nd ACM International Conference on Multimedia, 2024

  28. [36]

    Geometric transformer for fast and robust point cloud registration,

    Z. Qin, H. Yu, C. Wang, Y . Guo, Y . Peng, and K. Xu, “Geometric transformer for fast and robust point cloud registration,” in CVPR, 2022

  29. [37]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al. , “Dinov2: Learning robust visual features without supervision,” arXiv preprint arXiv:2304.07193, 2023

  30. [38]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., “Learning transferable visual models from natural language supervision,” in ICML, 2021

  31. [39]

    Covis-net: A cooperative visual spatial foundation model for multi-robot applications,

    J. Blumenkamp, S. Morad, J. Gielis, and A. Prorok, “Covis-net: A cooperative visual spatial foundation model for multi-robot applications,” CoRL, 2025

  32. [40]

    Deep graph matching consensus,

    M. Fey, J. E. Lenssen, C. Morris, J. Masci, and N. M. Kriege, “Deep graph matching consensus,” arXiv preprint arXiv:2001.09621 , 2020

  33. [41]

    The hungarian method for the assignment problem,

    H. W. Kuhn, “The hungarian method for the assignment problem,” Naval research logistics quarterly , vol. 2, pp. 83–97, 1955

  34. [42]

    Carla: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “Carla: An open urban driving simulator,” in CoRL, 2017

  35. [43]

    Sumo (simulation of urban mobility)-an open-source traffic simulation,

    D. Krajzewicz, G. Hertkorn, C. R ¨ossel, and P. Wagner, “Sumo (simulation of urban mobility)-an open-source traffic simulation,” in MESM, 2002

  36. [44]

    Multiagent multitraversal multimodal self-driving: Open mars dataset,

    Y . Li, Z. Li, N. Chen, M. Gong, Z. Lyu, Z. Wang, P. Jiang, and C. Feng, “Multiagent multitraversal multimodal self-driving: Open mars dataset,” in CVPR, 2024

  37. [45]

    ultralytics/yolov5: v6. 1-tensorrt, tensorflow edge tpu and openvino export and inference,

    G. Jocher, A. Chaurasia, A. Stoken, J. Borovec, Y . Kwon, J. Fang, K. Michael, D. Montes, J. Nadar, P. Skalski, et al., “ultralytics/yolov5: v6. 1-tensorrt, tensorflow edge tpu and openvino export and inference,” Zenodo, 2022

  38. [46]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,” arXiv:2406.09414, 2024

  39. [47]

    Fast graph representation learning with PyTorch Geometric,

    M. Fey and J. E. Lenssen, “Fast graph representation learning with PyTorch Geometric,” in ICLR Workshop on Representation Learning on Graphs and Manifolds , 2019

  40. [48]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

  41. [49]

    Splinecnn: Fast geometric deep learning with continuous b-spline kernels,

    M. Fey, J. E. Lenssen, F. Weichert, and H. M ¨uller, “Splinecnn: Fast geometric deep learning with continuous b-spline kernels,” in CVPR, 2018

  42. [50]

    Bayesian deep graph matching for correspon- dence identification in collaborative perception,

    P. Gao and H. Zhang, “Bayesian deep graph matching for correspon- dence identification in collaborative perception,” in RSS, 2021

  43. [51]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,

    M. FISCHLER AND, “Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,” Commun. ACM, p. 381–395, June 1981

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.