Pith. sign in

REVIEW 4 major objections 6 minor 34 references

A roadside 3D detector can infer and correct its own camera pose errors—near and far get different fixes—using only the image itself.

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 →

RECO adds learnable near/far 6-DoF pose corrections to roadside BEV detectors, smoothly blended by a sigmoid gate, improving 3D detection under camera extrinsic jitter and drift.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A plausible range-adaptive compensation design for roadside BEV detection, but the persistent-drift story is undercut by an ablation that suggests the pose head may be learning a degenerate, shift-invariant correction instead of actually estimating the injected perturbation. the 4 major comments →

arxiv 2607.20947 v1 pith:VNX345HJ submitted 2026-07-23 cs.CV

RECO: Region-Aware Compensation for Extrinsic Perturbations in Roadside 3D Detection

classification cs.CV
keywords roadside 3D object detectioncamera extrinsic calibrationextrinsic perturbationbird's-eye viewregion-aware compensation6-DoF pose estimationsoft gatingreprojection loss
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

The paper targets a practical failure mode in roadside 3D detection: cameras mounted on infrastructure slowly drift out of calibration, and even tiny pose changes get amplified by perspective projection into large bird's-eye-view errors, especially at long range. RECO's claim is that a model can recover the unknown pose error from the image itself by splitting the scene into near and far ranges and predicting separate six-degree-of-freedom corrections for each, then blending them with a smooth gate. A reprojection loss, comparing 3D boxes projected through the corrected geometry against 2D annotations, supervises the corrections jointly with the detection loss. The paper reports consistent gains over prior BEV detectors under both random jitter and persistent mean-shifted drift on two roadside benchmarks. The significance, if true: infrastructure cameras can self-heal their calibration online, without new hardware or manual re-calibration.

Core claim

The central claim is that a single global pose correction cannot absorb camera extrinsic perturbations in roadside BEV detection, because the error's character changes with range: near-range misalignment is dominated by translation bias, far-range by rotation sensitivity, and supervision signals are unevenly distributed across distance. RECO therefore predicts a learnable boundary and two SE(3) offset corrections—one for near, one for far—and uses a sigmoid gate to smoothly interpolate their projection geometries, preserving a continuous BEV sampling field. An auxiliary reprojection loss supervises the offsets by penalizing the L2 discrepancy between 2D boxes obtained by projecting 3D ground

What carries the argument

The mechanism is piecewise 6-DoF extrinsic compensation with differentiable soft gating. From image features and the nominal extrinsic, a region-aware module outputs a scalar range boundary b and two pose offsets, converted to SE(3) corrections ΔTn and ΔTf. Each correction is left-composed with the extrinsic to produce two candidate projection geometries; for each BEV cell at distance r, the sigmoid gate g = σ((r−b)/τ) blends them into a single continuous geometry field. The auxiliary reprojection loss drives this by comparing projected 3D box corners to 2D box annotations, so the network learns to cancel the geometric misalignment rather than merely disregard it.

Load-bearing premise

The load-bearing premise is that the visual features contain enough information to identify a nonzero persistent pose offset, even though the training noise is zero-mean; if the network only learns to output the average (zero) correction, the persistent-drift results collapse.

What would settle it

Train under zero-mean jitter and evaluate at a single fixed yaw offset of +3 degrees (well beyond the training scale). If RECO's predicted corrections stay near zero and AP falls to the no-compensation baseline, the persistent-drift generalization claim is unsupported. A second check: feed clean, unperturbed images and inspect the predicted offsets—if they are consistently zero, the model has not learned to extract pose from content.

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

If this is right

  • Infrastructure cameras can recover from slow drift (thermal expansion, structural deformation, remounting) without manual re-calibration, as long as image content constrains the pose.
  • Training on zero-mean jitter with region-aware compensation generalizes to nonzero mean offsets, so calibration errors do not need to be present during training to be fixed at test time.
  • Because compensation happens inside the projection operator, existing BEV detection heads can be reused unchanged.
  • The same trained model handles pitch, roll, and combined perturbations at test time without retraining, per the paper's rotation experiments.
  • The learnable boundary means the near/far partition adapts per scene rather than being hard-coded.

Where Pith is reading between the lines

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

  • The range-aware correction idea is likely transferable to other range-dependent projection errors—lens distortion, rolling-shutter artifacts, or per-pixel depth bias—where a single global warp is also a poor fit.
  • The reprojection-loss design suggests a self-supervised deployment recipe: use 2D boxes from an off-the-shelf 2D detector at a new site to fine-tune the compensation branch without 3D labels.
  • The method could be extended from a single camera to a multi-camera infrastructure network by predicting one boundary per camera and enforcing consistency in overlapping regions, a direction the paper names as future work.
  • An online temporal filter on predicted offsets could turn RECO into a continuous calibration monitor, flagging when drift exceeds a threshold rather than only correcting it per frame.
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 / 6 minor

Summary. The paper proposes RECO, a region-aware extrinsic compensation module for roadside monocular 3D detection. Given a single image and nominal camera extrinsics, a region-aware network predicts near- and far-range 6-DoF SE(3) offsets and a range boundary. A differentiable sigmoid gate blends the two compensated projection geometries to produce a continuous BEV sampling field (Eqs. 3-8). Training optimizes the standard 3D detection loss together with an auxiliary reprojection loss that compares 2D boxes projected from 3D ground truth against 2D annotations (Eqs. 11-12). The method is evaluated on DAIR-V2X-I and Rope3D under transient Gaussian extrinsic perturbations and under persistent mean-shifted deviations, reporting consistent improvements over existing BEV-based detectors. Ablations compare no compensation, global compensation, hard region assignment, and the proposed soft-gated compensation.

Significance. If the underlying inference mechanism works as claimed, RECO would be a meaningful contribution to robust infrastructure perception, offering a clean and continuous formulation for range-adaptive extrinsic correction (notably the soft-gated projection of Eqs. 7-8) and a sensible auxiliary reprojection loss. The reported gains in Tables 1 and 2 are substantial in many conditions. However, the central claim that the network actually estimates pose corrections is not directly supported: the Global_Comp ablation in Table 5 collapses at zero shift, no analysis of the predicted offsets is provided, and all experiments are single-run without code or variance. As it stands, the evidence is more consistent with a learned robust projection than with true pose inference, and the persistent-drift generalization claim rests on an unverified mechanism.

major comments (4)
  1. [Section 4.3, Table 5] The Global_Comp row is a critical diagnostic. At µ=0 under yaw, Global_Comp gives 33.24 AP vs 67.56 for No_Comp, and similarly for z-shift (23.27 vs 48.33). If the compensation head were recovering the injected pose perturbation, the optimal output at zero mean shift is near-identity; instead it converges to a severely detrimental constant offset. This is direct evidence that the reprojection/detection losses do not constrain predicted offsets to match true extrinsics. The flat persistent-drift curve of Soft_Comp in Table 3 (70.1 at µ=±1 vs 70.5 at µ=0) may therefore be due to the soft-gated geometry absorbing the shift in a model-specific way, not to pose inference. Please report the learned offsets and boundary as a function of the injected perturbation, their correlation with the true shift, and per-sample distribution; or reframe the contribution as a robust projection and support th
  2. [Tables 1-5] All experiments report a single run per condition with no variance, and no code/checkpoints are provided. Several claimed gains are small (e.g., Table 1 yaw car easy: 70.46 vs 69.69 for BEVHeight++, +0.77 AP), so without repeated trials or confidence intervals the SOTA statement is not statistically grounded. Please include mean/std over at least 3 seeds, or clearly explain the reporting protocol, and release code/checkpoints or the exact training details sufficient for reproduction.
  3. [Section 4.2, Table 3] The persistent-drift evidence covers only car AP at IoU=0.5 easy. The abstract's claim about maintaining strong performance under strict calibration uncertainty is further weakened by the z-shift results: at µ=-2, RECO AP drops from 72.6 (µ=0) to 31.1. Please provide all difficulty splits and other categories (pedestrian, cyclist) under persistent shifts, and discuss whether the observed asymmetry is consistent with the proposed range-compensation mechanism.
  4. [Section 4.2, first paragraph] The text states 'we compare RECO with ... calibration-free methods like CBR [2]', but Table 1 contains no CBR row. This is a direct inconsistency between text and table. Either include CBR results in Table 1, or move the statement to the supplementary material and reference it explicitly.
minor comments (6)
  1. [Eq. (2)] The notation ∆p is used both for a generic 6-DoF offset and for the near/far pair; please use ∆p_n and ∆p_f and define the stacking order.
  2. [Table 1 caption] Specify the units of the perturbation (degrees for yaw, meters for z) and clarify whether N(0,0.5) denotes standard deviation or variance.
  3. [Fig. 1 caption] The caption is garbled ('GT BoxPred BoxFar Near'); clean up the labels so the figure is self-contained.
  4. [Fig. 3 caption] The label '(c)' appears twice; likely one of the panels is mislabeled.
  5. [Eq. (10)] The scatter-pooling expression is hard to parse; write the aggregation as a sum over assigned cell indices with explicit normalization.
  6. [Section 3.3] The text mentions k-1 boundaries but the experiments use a single boundary; clarify that k=2 in this work and defer the general-k formulation to the supplementary material if needed.

Circularity Check

0 steps flagged

No significant circularity: RECO's pose compensation is supervised by external 2D annotations and independent benchmarks; the BEVSpread self-citation is not load-bearing.

full rationale

The central derivation chain is self-contained with respect to circularity. The predicted 6-DoF offsets and range boundary are supervised by a detection loss and by an auxiliary reprojection loss (Eq. 11) that compares 2D boxes projected from 3D ground truth against external 2D annotations; the combined objective is L = L_det + λ_rep L_rep (Eq. 12). These supervision signals are not defined in terms of the predicted offsets or boundary, so the predictions are not equivalent to their inputs by construction. The boundary head in Eq. 3 uses a fixed prior with a learned residual, and the soft-gate blend in Eqs. 7-8 combines geometries from the predicted offsets without reintroducing the targets as inputs. The benchmarks and baselines are independently published; one current author is also an author of the BEVSpread baseline, but that citation is used only for comparison and does not provide a load-bearing premise for RECO's design. The Table 5 Global_Comp collapse at µ=0 is a possible internal-validity weakness regarding whether the learned compensation corresponds to true pose errors, but it is an empirical correctness concern, not a circularity in the derivation. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no external result is replaced by a self-citation chain.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on the standard projection model plus a small set of design choices and one implicit identifiability assumption. The free hyperparameters (boundary prior, tau, lambda_rep, region count) are not swept, and the identifiability assumption is the main burden.

free parameters (5)
  • range boundary prior b-tilde = 20 m (default for near/far split; [20,40,...] for k>2)
    Hand-set before training; the boundary head adds a learned residual b = b_tilde + tanh(U_b Z). The choice affects which BEV cells use the near vs far correction.
  • soft gate temperature tau = 0.5
    Set by hand; controls the transition width between near and far geometries. No sensitivity analysis is provided.
  • reprojection loss weight lambda_rep = 0.1
    Set by hand to balance L_det and L_rep; no sensitivity analysis is provided.
  • perturbation scale sigma = 0.5 (N(0,0.5), degrees or meters)
    Training noise scale used to simulate transient perturbations; all methods are trained under the same protocol, but the chosen regime determines the scope of the robustness claims.
  • number of regions k = 2 (near/far)
    The text mentions generalization to k segments, but experiments use only two regions; the region count is not varied.
axioms (5)
  • standard math Pinhole camera model with known intrinsics K correctly maps 3D points to pixels (Eq. 5-6).
    Standard projective geometry, used to build the BEV sampling field and the reprojection loss.
  • domain assumption Range-dependent error dominance: near-range misalignment is dominated by translation, far-range by rotation (Sec. 3.3, citing [20]).
    This motivates the piecewise near/far correction; if error structure is not actually range-dominant in the tested datasets, the method loses its main justification.
  • ad hoc to paper Visual features contain enough information to estimate the current extrinsic perturbation (Sec. 3.3/4.2, implicit).
    The network must infer the pose offset from a single image, including unseen persistent shifts. The paper does not state or analyze this identifiability requirement.
  • domain assumption 2D box annotations are accurate projections of 3D ground-truth boxes, so minimizing L_rep drives Delta_p toward the true pose error (Eq. 11).
    The reprojection loss is the main supervisory signal for the compensation module; large 2D annotation noise would bias the learned offsets.
  • domain assumption Linear interpolation of pixel-coordinate fields G_n and G_f in Eq. 8 yields a sufficiently accurate compensated geometry.
    Soft blending of two projection grids is an approximation to the true range-dependent transform; its accuracy depends on the transition width and the smoothness of the error field.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of RECO: Region-Aware Compensation for Extrinsic Perturbations in Roadside 3D Detection." pith.science (2026). https://pith.science/paper/VNX345HJ

@misc{pith2026260720947,
  author       = {Pith},
  title        = {Pith review of: RECO: Region-Aware Compensation for Extrinsic Perturbations in Roadside 3D Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VNX345HJ}},
  note         = {Machine review of arXiv:2607.20947}
}
Share X Bluesky LinkedIn Reddit HN
abstract

In intelligent transportation systems, roadside 3D object detection provides wide-area perception crucial for traffic understanding, cooperative early warning, and safe autonomous driving. However, existing methods suffer from high sensitivity to camera extrinsics; even slight deviations (whether manifesting as transient jitter or persistent drift) can be significantly amplified by projective geometry. This cascade results in severe feature misalignment and degraded localization. To mitigate this limitation, we propose RECO, a region-aware extrinsic compensation framework that corrects extrinsics using piecewise 6-DoF pose offsets. RECO predicts a learnable range boundary to partition the scene into near and far regions, estimating region-specific pose corrections. A differentiable sigmoid gate then smoothly blends the two compensated geometries to preserve continuous BEV sampling and facilitate stable optimization. To supervise the refinement of extrinsics, we introduce an auxiliary reprojection loss that compares 2D bounding boxes projected from 3D ground truth against 2D annotations, optimizing it jointly with the standard detection objective. Extensive experiments on the DAIR-V2X-I and Rope3D benchmarks under extrinsic perturbations demonstrate consistent improvements over state-of-the-art baselines across both yaw and $z$-axis deviations. RECO also generalizes from transient perturbations to persistent shifts, maintaining highly competitive performance under strict calibration uncertainty.

Figures

Figures reproduced from arXiv: 2607.20947 by Junsheng Du, Yuhuan Lu, Zhaocheng He.

Figure 1
Figure 1. Figure 1: Illustration of camera pose deviations and their impact on roadside 3D detec￾tion. (a) Persistent pose drift caused by long-term structural changes. (b) Transient pose perturbations induced by external disturbances (e.g., vibrations) that may re￾bound to the nominal pose. (c) Using a fixed, pre-calibrated extrinsic pose under drift causes misaligned 3D predictions, with errors typically scaling proportiona… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of RECO, a region-aware extrinsic compensation framework for in￾frastructure monocular 3D object detection. The region-aware module predicts near/far 6-DoF pose offsets ∆p and a range boundary b, which are used by the compensation projection to construct BEV features for the detector head. homogeneous transform as \mathbf {T} = \begin {bmatrix} \mathbf {R} & t \\ 0^\top & 1 \end {bmatrix} \in \mat… view at source ↗
Figure 3
Figure 3. Figure 3: Motivation and design for our region-aware extrinsic compensation. (a) Range￾dependent effect of camera extrinsic perturbations: near-range misalignment is domi￾nated by translation bias, while far-range misalignment is more sensitive to rotation jitter (the blue box denotes the ground-truth, while red denotes the prediction). (b) A single global 6-DoF correction cannot simultaneously fit near and far rang… view at source ↗
Figure 4
Figure 4. Figure 4: Motivation for soft gating in range-aware extrinsic compensation. (a) Assigning piecewise offsets produces a piecewise projection, causing a discontinuous mapping at the boundary b (non-smooth BEV feature sampling). (b) A differentiable soft gate smoothly blends near/far geometries, yielding a continuous projection field and stable optimization. Here, G denotes the original projection space. G′ and G′′ ill… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison under extrinsic perturbations. Top: a representative roadside image (left) and the corresponding ground-truth annotations (right). Bot￾tom: BEV visualizations for different methods (rows). (a) Missed and false detections: red crosses denote missed ground-truth objects and unmatched predictions, while dots indicate ground-truth and matched predictions. (b) Distance error map for match… view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of BEV grid displacement and compensation under yaw N (0, 0.5). (a) Displacement magnitude after injecting extrinsic noise, measured as ||Gnoisy − Gbase|| at each BEV cell. (b) The compensation applied by our method, ||Gcomp − Gnoisy||, shows how the geometry field is corrected across the grid. (c) Residual dis￾placement after compensation, ||Gcomp−Gbase||. The overlaid polylines are iso-cont… 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

34 extracted references · 3 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition

    Aoki, Y., Goforth, H., Srivatsan, R.A., Lucey, S.: Pointnetlk: Robust & efficient point cloud registration using pointnet. In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition. pp. 7163–7172 (2019)

  2. [2]

    In: 2023 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS)

    Fan, S., Wang, Z., Huo, X., Wang, Y., Liu, J.: Calibration-free bev represen- tation for infrastructure perception. In: 2023 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS). pp. 9008–9013 (2023).https: //doi.org/10.1109/IROS55552.2023.10341916 16 J. Du et al

  3. [3]

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

    Fu, K., Liu, S., Luo, X., Wang, M.: Robust point cloud registration framework based on deep graph matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8893–8902 (2021)

  4. [4]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

  5. [5]

    arxiv 2022

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

  6. [6]

    arXiv preprint arXiv:2112.11790 (2021)

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

  7. [7]

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

    Huang, S., Gojcic, Z., Usvyatsov, M., Wieser, A., Schindler, K.: Predator: Reg- istration of 3d point clouds with low overlap. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 4267–4276 (2021)

  8. [8]

    In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Iyer, G., Ram, R.K., Murthy, J.K., Krishna, K.M.: Calibnet: Geometrically su- pervised extrinsic calibration using 3d spatial transformer networks. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1110–1117. IEEE (2018)

  9. [9]

    Iscience27(5) (2024)

    Ji, Y., Zhou, Z., Yang, Z., Huang, Y., Zhang, Y., Zhang, W., Xiong, L., Yu, Z.: To- ward autonomous vehicles: A survey on cooperative vehicle-infrastructure system. Iscience27(5) (2024)

  10. [10]

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

    Li, J., Lee, G.H.: Deepi2p: Image-to-point cloud registration via deep classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15960–15969 (2021)

  11. [11]

    In: Proceedings of the AAAI conference on artificial intelligence

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

  12. [12]

    IEEE Transactions on Pattern Analysis and Machine Intelligence47(3), 2020–2036 (2024)

    Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Yu, Q., Dai, J.: Bevformer: learningbird’s-eye-viewrepresentationfromlidar-cameraviaspatiotemporaltrans- formers. IEEE Transactions on Pattern Analysis and Machine Intelligence47(3), 2020–2036 (2024)

  13. [13]

    arXiv preprint arXiv:1711.05101 (2017)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

  14. [14]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Lv, X., Wang, B., Dou, Z., Ye, D., Wang, S.: Lccnet: Lidar and camera self- calibration using cost volume network. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 2894–2901 (2021)

  15. [15]

    Ma, Y., Li, Y., Hua, W., Kong, S.: Roadside monocular 3d detection prompted by 2ddetection.In:ProceedingsoftheIEEE/CVFWinterConferenceonApplications of Computer Vision (WACV) (2026)

  16. [16]

    In: European conference on computer vision

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

  17. [17]

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

    Qin, Z., Yu, H., Wang, C., Guo, Y., Peng, Y., Xu, K.: Geometric transformer for fast and robust point cloud registration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11143–11152 (2022)

  18. [18]

    IEEE Transactions on Circuits and Systems for Video Technology33(3), 1198–1208 (2022)

    Ren, S., Zeng, Y., Hou, J., Chen, X.: Corri2p: Deep image-to-point cloud regis- tration via dense correspondence. IEEE Transactions on Circuits and Systems for Video Technology33(3), 1198–1208 (2022)

  19. [19]

    In: 2017 IEEE intelligent vehicles symposium (IV)

    Schneider, N., Piewak, F., Stiller, C., Franke, U.: Regnet: Multimodal sensor regis- tration using deep neural networks. In: 2017 IEEE intelligent vehicles symposium (IV). pp. 1803–1810. IEEE (2017) RECO: Region-Aware Extrinsic Compensation for Roadside 3D Detection 17

  20. [20]

    National Research Council of Canada (2002)

    Scott, W.R., Roth, G., Rivest, J.F.: Pose error effects on range sensing. National Research Council of Canada (2002)

  21. [21]

    IEEE Transactions on Image Processing (2024)

    Shi, H., Pang, C., Zhang, J., Yang, K., Wu, Y., Ni, H., Lin, Y., Stiefelhagen, R., Wang, K.: Cobev: Elevating roadside 3d object detection with depth and height complementarity. IEEE Transactions on Image Processing (2024)

  22. [22]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Vuong, K., Tamburo, R., Narasimhan, S.G.: Toward planet-wide traffic camera calibration. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 8553–8562 (2024)

  23. [23]

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

    Wang, W., Lu, Y., Zheng, G., Zhan, S., Ye, X., Tan, Z., Wang, J., Wang, G., Li, X.: Bevspread: Spread voxel pooling for bird’s-eye-view representation in vision- based roadside 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14718–14727 (2024)

  24. [24]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Wang, Y., Solomon, J.M.: Deep closest point: Learning representations for point cloud registration. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3523–3532 (2019)

  25. [25]

    IEEE Geo- science and Remote Sensing Letters21, 1–5 (2024)

    Wu, J., Yuan, M., Wang, T., Jia, X., Yan, D.M.: Heightformer: Single-imagery height estimation transformer with bilateral feature pyramid fusion. IEEE Geo- science and Remote Sensing Letters21, 1–5 (2024)

  26. [26]

    In: 2025 IEEE/CVF Winter Confer- ence on Applications of Computer Vision (WACV)

    Xia, Z.X., Fadadu, S., Shi, Y., Foucard, L.: Robust long-range perception against sensor misalignment in autonomous vehicles. In: 2025 IEEE/CVF Winter Confer- ence on Applications of Computer Vision (WACV). pp. 5761–5770. IEEE (2025)

  27. [27]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA)

    Xiao, Y., Li, Y., Meng, C., Li, X., Ji, J., Zhang, Y.: Calibformer: A transformer- based automatic lidar-camera calibration network. In: 2024 IEEE International Conference on Robotics and Automation (ICRA). pp. 16714–16720. IEEE (2024)

  28. [28]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    Yang, L., Tang, T., Li, J., Yuan, K., Wu, K., Chen, P., Wang, L., Huang, Y., Li, L., Zhang, X., et al.: Bevheight++: Toward robust visual centric 3d object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

  29. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Yang, L., Yu, K., Tang, T., Li, J., Yuan, K., Wang, L., Zhang, X., Chen, P.: Bevheight: A robust framework for vision-based roadside 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. pp. 21611–21620 (2023)

  30. [30]

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

    Ye, X., Shu, M., Li, H., Shi, Y., Li, Y., Wang, G., Tan, X., Ding, E.: Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21341–21350 (2022)

  31. [31]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Yew, Z.J., Lee, G.H.: Rpm-net: Robust point matching using learned features. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 11824–11833 (2020)

  32. [32]

    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)

  33. [33]

    Sensors25(13) (2025), https://www.mdpi.com/1424-8220/25/13/3919

    Zhang, W., Gao, Y., Jiang, Z., Mao, R., Zhou, S.: Calibration-free roadside bev perception with v2x-enabled vehicle position assistance. Sensors25(13) (2025), https://www.mdpi.com/1424-8220/25/13/3919

  34. [34]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhu, Y., Wang, Z., Wang, Y.: Mamv2xcalib: V2x-based target-less infrastructure camera calibration with state space model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 26696–26705 (2025)

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