Pith. sign in

REVIEW 3 major objections 5 minor 50 references

RC-AutoCalib: An End-to-End Radar-Camera Automatic Calibration Network

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read RC-AutoCalib claims the first end-to-end online radar-camera calibration, reaching 0.427° rotation and 9.498 cm translation error on nuScenes.

desk verdict Useful end-to-end radar-camera calibration network with real novelty in dual-perspective matching, but the headline comparative claim is untested because Wise et al., the strongest prior radar-camera method, never appears in any comparison table. read the letter →

arxiv 2505.22427 v1 pith:JTBOO5BW submitted 2025-05-28 cs.CV

classification cs.CV
keywords radar-cameracalibrationonlineextrinsicbird's-eyeviewcross-attentionfeaturematchingsupervisionnuScenesmulti-modalfusion
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

Radar and camera are the cheap, weather-robust sensor pair for driver assistance, but radar points are so sparse and their heights so unreliable that online recalibration has lagged LiDAR-camera calibration. RC-AutoCalib attacks this by representing the scene twice: a frontal view rich in semantics but sensitive to height errors, and a bird's-eye view that is stable against height uncertainty, with a learned selection mechanism fusing only the reliable parts. Feature matching between radar and camera is supervised explicitly, and a Noise-Resistant Matcher uses LiDAR points to label which radar points are trustworthy before they seed the matching loss. On nuScenes the method reports 0.427° mean rotation error and 9.498 cm mean translation error, beating prior radar-camera auto-calibration and LiDAR-camera calibration methods. If this holds, sensors that drift over time could be recalibrated continuously without calibration targets, manual effort, or stopping the vehicle.

What carries the argument

The load-bearing element is the Noise-Resistant Matcher paired with Explicit Feature Matching Supervision. For every radar point the matcher builds an adaptive 3D bounding box whose height, width, and depth are derived from the radar's elevation-ambiguity geometry: $\delta$ is the allowed height error, $\Delta s$ an inter-sensor error offset, and $\Delta x,\Delta z$ the ground-plane shifts implied by an unknown elevation angle. A radar point is kept for supervision only if at least $\tau=3$ LiDAR points fall inside that box, which is how the network acquires clean positive matches despite noisy heights. Those matches supervise an assignment matrix $P$ (computed through a softmax similarity structure) with a log-likelihood loss, forcing cross-modal attention to concentrate on true correspondences rather than spurious zero-value correlations. The Dual-Perspective representation and Selective Fusion are supporting machinery: they make sure the network can fall back on height-robust bird's-eye features when frontal height information is unreliable.

What would settle it

Take a radar-camera pair with known ground-truth extrinsics whose radar elevation-ambiguity violates the nuScenes-tuned box (for example, a radar with a much wider vertical field of view or longer range), run the published pipeline with the same $\delta=1$, $\Delta s=0.5$, $\tau=3$ settings, and see whether rotation/translation errors rise toward or above the LiDAR-camera baselines; a sharp rise would show the noise-resistance assumption does not transfer. A second, cleaner test is to remove the LiDAR-supervised Noise-Resistant Matcher from training and measure whether the translation error stays below the reported 9.498 cm.

Watch

Extended reading notes

Core claim

The paper's central claim is that 6-DoF radar-camera extrinsic calibration can be solved end-to-end despite radar's extreme sparsity and elevation ambiguity, by making the network choose between two complementary views and by directly teaching it which point pairs really correspond. Concretely, RC-AutoCalib converts image and radar data into frontal depth maps and bird's-eye maps, matches features in each view with a Multi-Modal Cross-Attention mechanism, fuses the views with a Selective Fusion module, and regresses rotation and translation with an LSTM-based iterative head. The decisive ingredient is explicit matching supervision: the network predicts an assignment between radar and image features and is trained against matches generated from ground-truth calibration, after a Noise-Resistant Matcher removes radar points whose height uncertainty makes them unreliable. The reported outcome is a mean rotation error of 0.427° and a mean translation error of 9.498 cm on nuScenes, with ablations attributing the largest translation gains to the explicit matching supervision and the noise filter.

Load-bearing premise

The load-bearing premise is that the hand-set reliability rule—a radar point is trustworthy only if at least three LiDAR points fall inside a height-error box built from $\delta=1$ m, $\Delta s=0.5$ m, and $\tau=3$—correctly separates good from noisy radar training signal, and that this rule transfers to radar sensors and environments beyond nuScenes.

Editorial extensions

If this is right

  • Radar-camera systems can be recalibrated during normal operation, covering both rotation and translation, so mounting vibration or thermal drift no longer requires a workshop visit.
  • The explicit matching supervision is what unlocks translation accuracy: the ablation shows a 23.4% translation-error reduction when it is added, and a further drop when the Noise-Resistant Matcher filters unreliable points.
  • The same architecture transfers to LiDAR-camera calibration by removing the radar-specific matcher, and the supplementary experiments report it beating prior LiDAR-camera calibrators on both nuScenes and KITTI.
  • Models trained on nuScenes transfer to the aiMotive urban and rain scenarios with lower rotation error than all compared baselines, suggesting the learned correspondences are not dataset-specific.
  • The reported errors establish a public benchmark for radar-camera auto-calibration that later methods can be measured against.

Reading between the lines

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

  • The training pipeline depends on LiDAR as a teacher to decide which radar points are matchable; a fair test for deployment without LiDAR would be whether the same thresholds transfer, or whether a reliability prior learned from radar alone can substitute.
  • Because the depth branch relies on monocular metric depth estimation (DepthAnything plus ZoeDepth), the calibration ceiling is tied to depth-estimation quality; perturbing estimated depth with controlled synthetic noise would reveal how much of the reported accuracy rests on the depth network.
  • The noise-box parameters ($\delta=1$, $\Delta s=0.5$, $\tau=3$) are tuned for nuScenes' radar and LiDAR geometry; radar types with different elevation spread—e.g., long-range highway radars—would likely need re-tuning, offering a direct generalization test.
  • The same dual-perspective, selective-fusion recipe may apply to other sparse and height-ambiguous depth sensors such as sonar or low-cost LiDAR, since the architecture's core assumption is only that one view is semantically rich but noisy in height while another is height-robust.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript proposes RC-AutoCalib, an end-to-end network that estimates 6-DoF extrinsic calibration between a 3D radar and a camera from a single RGB image and radar point cloud. The method projects both modalities into frontal and bird's-eye views, applies cross-attention feature matching, fuses the two views with a selective mechanism, and regresses the calibration with an LSTM-based iterative head. A Noise-Resistant Matcher uses LiDAR points during training to filter unreliable radar points before supervising the matching loss. Experiments on nuScenes report mean rotation error 0.427 degrees and translation error 9.498 cm under small mis-calibration, with ablations, cross-dataset evaluation on aiMotive, and extension to LiDAR-camera calibration.

Significance. If the reported results hold, RC-AutoCalib is a practical contribution: it addresses sparse and height-noisy radar data with a reasonable dual-view representation, and it is one of the few deep-learning targetless methods for 6-DoF radar-camera calibration. The paper ships code, includes ablations for each module, and reports cross-dataset and downstream-task evaluations, all of which are strengths. However, the headline comparative claim is not yet established because the strongest targetless 3D radar-camera baselines (Wise et al.) are absent from all comparison tables, and the cross-dataset generalization evidence is limited to rotation.

major comments (3)
  1. [Section 2.2, Table 1] The paper identifies Wise et al. [39,40] as targetless 3D radar-to-camera extrinsic calibration methods that estimate translation, yet neither method appears in Table 1 or in the supplementary cross-dataset table. The only radar-camera baseline is Scholler et al. [31], which the authors themselves describe as rotation-only. Consequently, the abstract and conclusion claim that RC-AutoCalib 'significantly outperforms previous radar-camera auto-calibration methods' is not actually tested against the strongest relevant method. Please add Wise et al. on the same nuScenes split and mis-calibration ranges, or explicitly restrict the claim to learning-based radar-camera methods and discuss the missing comparison.
  2. [Section 3.3 and Supplementary D] The Noise-Resistant Matcher depends on three hand-set parameters (delta=1, Delta_s=0.5, tau=3) and on a LiDAR-based reliability criterion, but the paper provides no sensitivity analysis for these values and no evidence that the criterion transfers across radar sensors. Since the aiMotive cross-dataset experiment reports rotation errors only, the paper does not demonstrate that the matching supervision, and hence the learned calibration, transfers to radar types or environments unseen in nuScenes. I do not view the LiDAR-based label filtering as circular; the issue is that the thresholds are load-bearing for the generalization claim and are not tested.
  3. [Abstract and Introduction] The claim of being the 'first online automatic geometric calibration method for radar and camera systems' is contradicted by the authors' own related-work discussion, which lists Persic et al. [28] and Wise et al. [39,40] as earlier targetless or online radar-camera calibration methods. If 'first' is intended only for end-to-end deep-learning 6-DoF methods, the sentence should say so explicitly.
minor comments (5)
  1. [Eq. (12)] The variable Delta_y appears in Eq. (12) without a definition; the equality to 2(delta+Delta_s) suggests Delta_y is meant to be delta, and this should be clarified.
  2. [Sections 4.2 and 4.3] The text in Section 4.2 says 'As shown in Sec. 4.2' and Section 4.3 says 'In Sec. 4.3' when referring to the experimental results; these self-references should point to the relevant tables instead.
  3. [Supplementary Table 1] The aiMotive cross-dataset evaluation reports rotation errors only; reporting translation errors would substantially strengthen the generalization claim, since the main method claims 6-DoF calibration.
  4. [Table 1 caption] The caption should state explicitly that Scholler et al. [31] provides rotation-only results and that translation cells are therefore empty, to avoid apparent missing entries.
  5. [Abstract] The word 'groundbreaking' in the abstract is promotional; consider replacing it with a neutral description such as 'novel' or 'new'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the calibration errors are measured against held-out nuScenes ground truth, and the matching supervision is generated from that external ground truth plus LiDAR, not from the network's own outputs.

full rationale

I walked the derivation chain. The target calibration values come from nuScenes ground-truth extrinsic matrices; the network is trained with L_calib against T_gt and with L_matching whose true match matrix M is 'dynamically computed based on the true translation T_gt' (Sec. 3.6). This is standard supervised training, and test errors are computed against the same held-out ground truth, not against values fitted from the test set. The Noise-Resistant Matcher supervises matching using LiDAR points inside a 3D bounding box (Eqs. 12-14), an independent sensor modality; its hand-set thresholds delta=1, delta_s=0.5, tau=3 are assumptions, not fitted predictions. The paper contains no self-citation chain used as evidence: the cited prior methods [31,39,40] are external works, and no 'uniqueness theorem' from the authors is invoked. The claim of outperforming previous radar-camera auto-calibration methods is weakened by the absence of Wise et al. [39,40] from Tables 1 and the cross-dataset Table 1, and the 'first online automatic geometric calibration' claim is historically questionable given [39,40]; however, missing controls and priority overstatements are correctness risks, not circular derivation. No equation or reported number reduces to its own input by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method contributes a network architecture rather than a derivation. Its load-bearing components are supervised by ground-truth extrinsics and pretrained depth models. The hand-set parameters of the noise filter and loss weights are free parameters that affect results, selected by ablations rather than derived.

free parameters (6)
  • delta (allowable height error) = 1 (meter)
    Hand-set threshold for the noise-resistant matcher's bounding box height, controlling which LiDAR points count as neighbors (Eq. 12).
  • Delta_s (fixed sensor error offset) = 0.5 (meter)
    Hand-set constant added to bounding box dimensions to account for radar-LiDAR discrepancy (Section D).
  • tau (LiDAR neighbor threshold) = 3
    Radar points with fewer than tau LiDAR points inside the box are excluded from positive matches; chosen manually (Section D).
  • lambda (positive-negative balance) = 0.75
    Weight balancing positive and negative matching loss terms; selected via ablation (supplementary Table E.2).
  • beta (matching loss weight) = 0.1
    Weight of matching loss in total loss; selected via ablation (Table 4).
  • N (LSTM iteration steps) = 3
    Fixed number of autoregressive refinement iterations in the regression head.
assumptions (4)
  • domain assumption Radar elevation ambiguity and resulting x/z errors are correctly modeled by the equations from Singh et al. used to build the 3D bounding box.
    The noise-resistant matcher's box geometry rests on this prior noise model; if the model is inaccurate for nuScenes radar, filtering will mislabel training matches.
  • domain assumption Depth maps from DepthAnything and ZoeDepth provide metric depth accurate enough for cross-modal matching.
    The camera-side FV depth map and pseudo-BEV are generated from these pretrained models; calibration accuracy depends on their depth quality.
  • ad hoc to paper LiDAR points within the adaptive bounding box correctly identify reliable radar reflections.
    This heuristic is introduced by the paper and validated only indirectly through final calibration accuracy; no independent evidence is given that the thresholding is optimal.
  • domain assumption nuScenes provides accurate ground-truth extrinsic calibration for radar, camera, and LiDAR.
    All supervision and evaluation use these ground-truth transforms.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RC-AutoCalib: An End-to-End Radar-Camera Automatic Calibration Network." pith.science (2026). https://pith.science/paper/JTBOO5BW

@misc{pith2026250522427,
  author       = {Pith},
  title        = {Pith review of: RC-AutoCalib: An End-to-End Radar-Camera Automatic Calibration Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JTBOO5BW}},
  note         = {Machine review of arXiv:2505.22427}
}
read the original abstract

This paper presents a groundbreaking approach - the first online automatic geometric calibration method for radar and camera systems. Given the significant data sparsity and measurement uncertainty in radar height data, achieving automatic calibration during system operation has long been a challenge. To address the sparsity issue, we propose a Dual-Perspective representation that gathers features from both frontal and bird's-eye views. The frontal view contains rich but sensitive height information, whereas the bird's-eye view provides robust features against height uncertainty. We thereby propose a novel Selective Fusion Mechanism to identify and fuse reliable features from both perspectives, reducing the effect of height uncertainty. Moreover, for each view, we incorporate a Multi-Modal Cross-Attention Mechanism to explicitly find location correspondences through cross-modal matching. During the training phase, we also design a Noise-Resistant Matcher to provide better supervision and enhance the robustness of the matching mechanism against sparsity and height uncertainty. Our experimental results, tested on the nuScenes dataset, demonstrate that our method significantly outperforms previous radar-camera auto-calibration methods, as well as existing state-of-the-art LiDAR-camera calibration techniques, establishing a new benchmark for future research. The code is available at https://github.com/nycu-acm/RC-AutoCalib.

Figures

Figures reproduced from arXiv: 2505.22427 by the authors.

Figure 1
Figure 1. An overview of the proposed RC-AutoCalib method. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Challenges of 3D Millimeter-Wave Radar. (a) The green [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Our system flow for iterative online auto-calibration starts with the input image, point cloud, and initial calibration parameters [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of the proposed Feature Matching module. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Details of the Selective Fusion Mechanism [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Results visualization in the front view. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 1
Figure 1. Figure 1: Illustration of bounding box B. Suppose we consider only the y and z axes to calculate wB based on δ The cross-attention maps II←R, IR←I between radar and im￾age features will be computed according to the following equation: II←R = reshape( mmax j (Softmax(aIR)ij ),(h,…
Figure 2
Figure 2. Figure 2: Calibration results by projecting radar points onto the FV image [PITH_FULL_IMAGE:figures/full_fig_p011_2.png]
Figure 3
Figure 3. Figure 3: Examples of FV cross-attention maps highlight the important regions the model focuses on. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 40 canonical work pages

  1. [31]

    Targetless rotational auto-calibration of radar and camera for intelligent transportation systems

    Christoph Sch ¨oller, Maximilian Schnettler, Annkathrin Kr¨ammer, Gereon Hinz, Maida Bakovic, M ¨uge G ¨uzet, and Alois Knoll. Targetless rotational auto-calibration of radar and camera for intelligent transportation systems. In2019 IEEE Intelligent Transportation Systems Conference (ITSC), pages 3934–3941. IEEE, 2019. 1, 3, 7

  2. [28]

    Online multi-sensor calibration based on moving object tracking.Advanced Robotics, 35(3-4):130–140, 2021

    Juraj Per ˇsi´c, Luka Petrovi ´c, Ivan Markovi ´c, and Ivan Petrovi´c. Online multi-sensor calibration based on moving object tracking.Advanced Robotics, 35(3-4):130–140, 2021. 3

  3. [1]

    Layer normalization.arXiv preprint arXiv:1607.06450,

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hin- ton. Layer normalization.arXiv preprint arXiv:1607.06450,

  4. [2]

    Zoedepth: Zero-shot trans- fer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, 2023

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M ¨uller. Zoedepth: Zero-shot trans- fer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, 2023. 4

  5. [3]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020. 7

  6. [4]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 1

  7. [5]

    An ex- trinsic calibration tool for radar, camera and lidar

    Joris Domhof, Julian FP Kooij, and Dariu M Gavrila. An ex- trinsic calibration tool for radar, camera and lidar. In2019 In- ternational Conference on Robotics and Automation (ICRA), pages 8107–8113. IEEE, 2019. 1

  8. [6]

    Radar and vision sensors calibration for outdoor 3d reconstruction

    Ghina El Natour, Omar Ait Aider, Raphael Rouveure, Franc ¸ois Berry, and Patrice Faure. Radar and vision sensors calibration for outdoor 3d reconstruction. In2015 IEEE In- ternational Conference on Robotics and Automation (ICRA), pages 2084–2089. IEEE, 2015. 1, 3

Show all 50 references
  1. [7]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 4, 1

  2. [8]

    Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016. 1

  3. [9]

    Ehinger, and Tom Drummond

    Markus Hiller, Krista A. Ehinger, and Tom Drummond. Per- ceiving longer sequences with bi-directional cross-attention transformers.ArXiv, abs/2402.12138, 2024. 5

  4. [10]

    Batch normalization: Accelerating deep network training by reducing internal co- variate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal co- variate shift. InInternational conference on machine learn- ing, pages 448–456. pmlr, 2015. 6

  5. [11]

    Calibnet: Geometrically supervised extrinsic calibration using 3d spatial transformer networks

    Ganesh Iyer, R Karnik Ram, J Krishna Murthy, and K Mad- hava Krishna. Calibnet: Geometrically supervised extrinsic calibration using 3d spatial transformer networks. In2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1110–1117. IEEE, 2018. 1, 3, 4

  6. [12]

    Dxq-net: differentiable lidar-camera extrinsic calibration using quality-aware flow

    Xin Jing, Xiaqing Ding, Rong Xiong, Huanjun Deng, and Yue Wang. Dxq-net: differentiable lidar-camera extrinsic calibration using quality-aware flow. In2022 IEEE/RSJ In- ternational Conference on Intelligent Robots and Systems (IROS), pages 6235–6241. IEEE, 2022. 1

  7. [13]

    Data fusion of radar and image measurements for multi-object tracking via kalman filtering.Information Sciences, 278:641–652, 2014

    Du Yong Kim and Moongu Jeon. Data fusion of radar and image measurements for multi-object tracking via kalman filtering.Information Sciences, 278:641–652, 2014. 1, 3

  8. [14]

    Radar and vision sensor fusion for object detection in au- tonomous vehicle surroundings

    Jihun Kim, Dong Seog Han, and Benaoumeur Senouci. Radar and vision sensor fusion for object detection in au- tonomous vehicle surroundings. In2018 tenth international conference on ubiquitous and future networks (ICUFN), pages 76–78. IEEE, 2018. 3

  9. [15]

    Comparative analysis of radar-ir sensor fusion methods for object detection

    Taehwan Kim, Sungho Kim, Eunryung Lee, and Miryong Park. Comparative analysis of radar-ir sensor fusion methods for object detection. In2017 17th International Conference on Control, Automation and Systems (ICCAS), pages 1576–

  10. [16]

    Crn: Camera radar net for accurate, robust, efficient 3d perception

    Youngseok Kim, Juyeb Shin, Sanmin Kim, In-Jae Lee, Jun Won Choi, and Dongsuk Kum. Crn: Camera radar net for accurate, robust, efficient 3d perception. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 17615–17626, 2023. 3

  11. [17]

    Automatic online cal- ibration of cameras and lasers

    Jesse Levinson and Sebastian Thrun. Automatic online cal- ibration of cameras and lasers. InRobotics: science and systems. Citeseer, 2013. 3

  12. [18]

    Selec- tive kernel networks

    Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selec- tive kernel networks. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 510–519, 2019. 6

  13. [19]

    Automatic targetless lidar– camera calibration: a survey.Artificial Intelligence Review, 56(9):9949–9987, 2023

    Xingchen Li, Yuxuan Xiao, Beibei Wang, Haojie Ren, Yany- ong Zhang, and Jianmin Ji. Automatic targetless lidar– camera calibration: a survey.Artificial Intelligence Review, 56(9):9949–9987, 2023. 3

  14. [20]

    Bevdepth: Acquisition of reliable depth for multi-view 3d object detec- tion, 2022

    Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detec- tion, 2022. 1

  15. [21]

    Lightglue: Local feature matching at light speed

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Polle- feys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023. 5

  16. [22]

    Zhijian Liu, Haotian Tang, Sibo Zhu, and Song Han. Sema- lign: Annotation-free camera-lidar calibration with seman- tic alignment loss.2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8845–8851,

  17. [23]

    Radar- camera pixel depth association for depth completion

    Yunfei Long, Daniel Morris, Xiaoming Liu, Marcos Cas- tro, Punarjay Chakravarty, and Praveen Narayanan. Radar- camera pixel depth association for depth completion. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12507–12516, 2021. 1, 2

  18. [24]

    Lccnet: Lidar and camera self-calibration using cost volume network

    Xudong Lv, Boya Wang, Ziwen Dou, Dong Ye, and Shuo Wang. Lccnet: Lidar and camera self-calibration using cost volume network. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 2894–2901, 2021. 1, 2, 3, 4, 7

  19. [25]

    Cfnet: Lidar-camera registration using calibration flow network.Sensors, 21(23): 8112, 2021

    Xudong Lv, Shuo Wang, and Dong Ye. Cfnet: Lidar-camera registration using calibration flow network.Sensors, 21(23): 8112, 2021. 1

  20. [26]

    aimotive dataset: A multimodal dataset for robust au- tonomous driving with long-range perception.arXiv preprint arXiv:2211.09445, 2022

    Tam ´as Matuszka, Iv ´an Barton, ´Ad´am Butykai, P ´eter Ha- jas, D ´avid Kiss, Domonkos Kov ´acs, S ´andor Kuns ´agi- M´at´e, P ´eter Lengyel, G ´abor N ´emeth, Levente Pet ˝o, et al. aimotive dataset: A multimodal dataset for robust au- tonomous driving with long-range perce...

  21. [27]

    Automatic targetless extrinsic calibration of a 3d li- dar and camera by maximizing mutual information

    Gaurav Pandey, James McBride, Silvio Savarese, and Ryan Eustice. Automatic targetless extrinsic calibration of a 3d li- dar and camera by maximizing mutual information. InPro- ceedings of the AAAI conference on artificial intelligence, pages 2053–2059, 2012. 3

  22. [29]

    Long short-term memory based recurrent neural network architec- tures for large vocabulary speech recognition.arXiv preprint arXiv:1402.1128, 2014

    Has ¸im Sak, Andrew Senior, and Franc ¸oise Beaufays. Long short-term memory based recurrent neural network architec- tures for large vocabulary speech recognition.arXiv preprint arXiv:1402.1128, 2014. 3, 6

  23. [30]

    Regnet: Multimodal sensor registration using deep neural networks

    Nick Schneider, Florian Piewak, Christoph Stiller, and Uwe Franke. Regnet: Multimodal sensor registration using deep neural networks. In2017 IEEE intelligent vehicles sympo- sium (IV), pages 1803–1810. IEEE, 2017. 1, 3

  24. [32]

    Calnet: Lidar-camera online calibration with channel attention and liquid time- constant network

    Hongcheng Shang and Bin-Jie Hu. Calnet: Lidar-camera online calibration with channel attention and liquid time- constant network. In2022 26th International Conference on Pattern Recognition (ICPR), pages 5147–5154, 2022. 3

  25. [33]

    Calibrcnn: Calibrating camera and lidar by recurrent convolutional neu- ral network and geometric constraints

    Jieying Shi, Ziheng Zhu, Jianhua Zhang, Ruyu Liu, Zhen- hua Wang, Shengyong Chen, and Honghai Liu. Calibrcnn: Calibrating camera and lidar by recurrent convolutional neu- ral network and geometric constraints. In2020 IEEE/RSJ International Conference on Intelligent Robots and ...

  26. [34]

    Depth estimation from camera image and mmwave radar point cloud

    Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9275–92...

  27. [35]

    Obstacle detection using millimeter- wave radar and its visualization on image sequence

    Shigeki Sugimoto, Hayato Tateda, Hidekazu Takahashi, and Masatoshi Okutomi. Obstacle detection using millimeter- wave radar and its visualization on image sequence. InPro- ceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., pages 342–345. IE...

  28. [36]

    Motion-based calibration of multimodal sensor arrays

    Zachary Taylor and Juan Nieto. Motion-based calibration of multimodal sensor arrays. In2015 IEEE International Con- ference on Robotics and Automation (ICRA), pages 4843–

  29. [37]

    Fusionnet: Coarse-to-fine extrinsic calibration net- work of lidar and camera with hierarchical point-pixel fu- sion

    Guangming Wang, Jiahao Qiu, Yanfeng Guo, and Hesheng Wang. Fusionnet: Coarse-to-fine extrinsic calibration net- work of lidar and camera with hierarchical point-pixel fu- sion. In2022 International Conference on Robotics and Au- tomation (ICRA), pages 8964–8970. IEEE, 2022. 1, 2, 4

  30. [38]

    Integrating millimeter wave radar with a monocular vision sensor for on-road obstacle detection applications.Sensors, 11(9):8992–9008, 2011

    Tao Wang, Nanning Zheng, Jingmin Xin, and Zheng Ma. Integrating millimeter wave radar with a monocular vision sensor for on-road obstacle detection applications.Sensors, 11(9):8992–9008, 2011. 1, 3

  31. [39]

    A continuous-time ap- proach for 3d radar-to-camera extrinsic calibration

    Emmett Wise, Juraj Per ˇsi´c, Christopher Grebe, Ivan Petrovi´c, and Jonathan Kelly. A continuous-time ap- proach for 3d radar-to-camera extrinsic calibration. In2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13164–13170. IEEE, 2021. 3

  32. [40]

    Spa- tiotemporal calibration of 3-d millimetre-wavelength radar- camera pairs.IEEE Transactions on Robotics, 2023

    Emmett Wise, Qilong Cheng, and Jonathan Kelly. Spa- tiotemporal calibration of 3-d millimetre-wavelength radar- camera pairs.IEEE Transactions on Robotics, 2023. 3

  33. [41]

    Netcalib: A novel approach for lidar-camera auto- calibration based on deep learning

    Shan Wu, Amnir Hadachi, Damien Vivet, and Yadu Prab- hakar. Netcalib: A novel approach for lidar-camera auto- calibration based on deep learning. In2020 25th Inter- national Conference on Pattern Recognition (ICPR), pages 6648–6655. IEEE, 2021. 1, 2

  34. [42]

    This is the way: Sensors auto-calibration approach based on deep learning for self-driving cars.IEEE Sensors Journal, 21(24):27779–27788, 2021

    Shan Wu, Amnir Hadachi, Damien Vivet, and Yadu Prab- hakar. This is the way: Sensors auto-calibration approach based on deep learning for self-driving cars.IEEE Sensors Journal, 21(24):27779–27788, 2021. 1, 7

  35. [43]

    Sparse fuse dense: Towards high quality 3d detection with depth completion

    Xiaopei Wu, Liang Peng, Honghui Yang, Liang Xie, Chenxi Huang, Chengqi Deng, Haifeng Liu, and Deng Cai. Sparse fuse dense: Towards high quality 3d detection with depth completion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5418–5427,

  36. [44]

    Empirical evaluation of rectified activations in convolutional network

    Bing Xu, Naiyan Wang, Tianqi Chen, and Mu Li. Empirical evaluation of rectified activations in convolutional network. arXiv preprint arXiv:1505.00853, 2015. 5

  37. [45]

    Radarnet: Exploiting radar for robust per- ception of dynamic objects

    Bin Yang, Runsheng Guo, Ming Liang, Sergio Casas, and Raquel Urtasun. Radarnet: Exploiting radar for robust per- ception of dynamic objects. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16, pages 496–512. Sprin...

  38. [46]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024. 4

  39. [47]

    Pixel-level extrinsic self calibration of high resolution lidar and camera in targetless environments.IEEE Robotics and Automation Letters, 6(4):7517–7524, 2021

    Chongjian Yuan, Xiyuan Liu, Xiaoping Hong, and Fu Zhang. Pixel-level extrinsic self calibration of high resolution lidar and camera in targetless environments.IEEE Robotics and Automation Letters, 6(4):7517–7524, 2021. 1, 2, 3

  40. [48]

    Calibdnn: multimodal sensor calibration for perception us- ing deep neural networks

    Ganning Zhao, Jiesi Hu, Suya You, and C-C Jay Kuo. Calibdnn: multimodal sensor calibration for perception us- ing deep neural networks. InSignal Processing, Sen- sor/Information Fusion, and Target Recognition XXX, pages 324–335. SPIE, 2021. 2, 4, 3

  41. [49]

    Calibdepth: Uni- fying depth map representation for iterative lidar-camera on- line calibration

    Jiangtong Zhu, Jianru Xue, and Pu Zhang. Calibdepth: Uni- fying depth map representation for iterative lidar-camera on- line calibration. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 726–733. IEEE,

  42. [50]

    Perception-aware multi-sensor fusion for 3d lidar semantic segmentation

    Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuan- qing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16280–16290, 2021. 1

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.