Pith. sign in

REVIEW 4 major objections 7 minor 22 references

Multimodal HD Mapping for Intersections by Intelligent Roadside Units

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fusing camera imagery with LiDAR from roadside units improves intersection semantic HD map accuracy by 4 to 18 percentage points over either sensor alone.

desk verdict The RS-seq dataset is a genuine contribution, but the paper's headline fusion claim is contradicted by its own Table II and needs to be restricted to the ViT-Adapter setting. read the letter →

arxiv 2507.08903 v1 pith:XLBG3MWF submitted 2025-07-11 cs.RO cs.CV

Zhongzhang Chen , Miao Fan , Shengtong Xu , Mengmeng Yang , Kun Jiang , Xiangzeng Liu , Haoyi Xiong This is my paper
classification cs.ROcs.CV
keywords high-definitionsemanticmappingcamera-LiDARfusionintelligentroadsideunitsintersectionsegmentationvectorizedmapsRS-seqdatasetinfrastructure-assistedautonomousdriving
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

This paper tries to establish that HD semantic maps of intersections can be generated automatically from elevated roadside units by fusing camera images with LiDAR point clouds, and that the fused result is more accurate than either sensor alone. To test this, the paper builds RS-seq, a dataset of seven intersections with labels for lane dividers, pedestrian crossings, and stop lines in both camera images and LiDAR point clouds, and proposes a two-stage fusion baseline. Its main result is a mean IoU improvement of 4 percentage points over image-only segmentation and 18 percentage points over point-cloud-only segmentation. If correct, this gives infrastructure-assisted autonomous driving a practical path to keeping intersection maps current without repeated vehicle passes.

What carries the argument

The load-bearing machinery is a two-stage fusion pipeline built around the ground plane. Stage one performs data-level fusion: the RANSAC-extracted ground point cloud is projected through the camera's extrinsic and intrinsic parameters into image pixel coordinates, so each ground point inherits the class label of the image pixel it lands on; in parallel, the same ground points are gridded into small cells and rendered as an intensity image whose grey values are mean LiDAR reflectivity, which is then segmented by an image segmentation network. Stage two performs result-level fusion: each element's point clusters are denoised with statistical outlier removal, vectorized with alpha-shape edge extraction for polygons and least-squares fitting for lines, and the image-derived and LiDAR-derived vector maps are merged by set union. This lets the two modalities operate in their native representations and meet only at the final vector map, which is how the paper claims the complementarity is captured without forcing a shared feature space.

What would settle it

Run the released RS-seq data and code with a single strictly enforced synchronization threshold, say 20 ms, and independently re-calibrated camera-LiDAR extrinsics; if the multimodal mIoU advantage over the best unimodal baseline falls to near zero, the claimed complementarity would be an alignment artifact. A second check is to apply the same pipeline to a never-seen intersection from a different city with high-precision ground truth: sustained multi-point mIoU gains would corroborate the claim, while vanishing gains would indicate overfitting to the seven annotated intersections.

Watch

Extended reading notes

Core claim

The paper's central claim is that camera and LiDAR data collected from an elevated roadside unit are genuinely complementary for building semantic HD maps of intersections: the camera's texture and colour resolve lane dividers, while LiDAR's geometry resolves pedestrian crossings and stop lines, and a fusion of the two beats either sensor alone. To demonstrate this, the paper constructs RS-seq, a dataset of seven intersections with pixel- and point-level labels for lane dividers, pedestrian crossings, and stop lines, and proposes a two-stage fusion pipeline. In the first stage, ground LiDAR points are projected into the camera image so they inherit image-segmentation labels, while the ground point cloud is also gridded into an intensity image and segmented independently. In the second stage, the two modalities' segmented points are denoised, vectorized, and merged by taking the union of the vector maps. On the RS-seq test set the multimodal output reaches mIoU 0.474, which is 0.043 (4 percentage points) above image-only and 0.181 (18 percentage points) above point-cloud-only; the paper also reports that three established end-to-end vehicle-trained mapping models transfer poorly to the roadside perspective, which is the gap RS-seq is meant to fill.

Load-bearing premise

The whole comparison rests on camera and LiDAR frames being aligned precisely in space and time; the paper quotes two different synchronization tolerances and offsets the released coordinates, so if that alignment is wrong the fusion gains could be an artifact of the alignment rather than true sensor complementarity.

Editorial extensions

If this is right

  • If the 4 and 18 percentage-point gains are real, intersection HD maps can be updated from roadside feeds instead of dedicated mapping vehicles, lowering cost and speeding up updates.
  • Vehicle-trained end-to-end HD map models should not be expected to transfer directly to roadside data; RS-seq provides a benchmark for adapting them.
  • Accumulating more LiDAR frames improves mIoU from 0.268 at 1 frame to 0.498 at 100 frames, but computing time grows by over a hundred times, so deployments must balance map freshness against computational cost.
  • The proposed geometric fusion pipeline is a strong baseline for future learning-based roadside mapping, since it outperforms both unimodal inputs and direct transfer of established end-to-end baselines.
  • Transformer-based image segmentation models, such as the ViT-Adapter variant tested, achieve the best multimodal results on roadside data among the architectures compared.

Reading between the lines

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

  • The reported distance falloff implies that a single roadside unit will leave far-side map elements under-resolved; a second unit or a predictive completion module would be needed for full-intersection coverage.
  • The fusion recipe of borrowing image labels by projecting ground LiDAR points and then unioning vectorized outputs does not depend on the specific road elements tested, so it is a plausible candidate for labelling curbs, traffic signs, or lane endings once annotations exist.
  • The paper quotes two temporal synchronization tolerances, 50 ms and 0.02 s, so alignment quality should be treated as a controlled variable in follow-up work; if stricter synchronization erases the 4% gain, the fusion advantage may be largely a calibration advantage.
  • A test the paper does not run is a score-level late fusion that averages softmax maps before vectorization; this would isolate whether the measured gain comes from label transfer through projection or from the vectorization union step.
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

4 major / 7 minor

Summary. The paper proposes a camera-LiDAR fusion framework for high-definition semantic mapping of intersections using elevated intelligent roadside units (IRUs), and introduces RS-seq, a dataset of seven intersections with camera images, LiDAR point clouds, and vectorized map annotations for lane dividers, pedestrian crossings, and stop lines. The method combines data-level fusion (projecting point-cloud ground returns onto image segmentation results) and result-level fusion (vectorizing each modality's segmentation and taking the union). The authors evaluate with U-Net, PidNet, and ViT-Adapter under image-only, point-cloud-only, and multimodal settings, and report that multimodal fusion improves mIoU by 4% over image-only and 18% over point-cloud-only results. The central claim is that the multimodal approach consistently surpasses unimodal methods.

Significance. If the central claim held across models and configurations, the paper would provide a useful baseline for infrastructure-assisted HD mapping and a valuable new dataset. The RS-seq dataset appears to be a genuinely useful resource: it addresses a gap in roadside, intersection-focused multimodal datasets, and the authors state that they release both the dataset and source code, which supports reproducibility. The proposed fusion pipeline is simple and interpretable, and the comparison against HDMapNet, BEVFusion, and SuperFusion on roadside data is a useful reference point. However, the paper's headline claim is not supported by its own Table II for the CNN backbones tested, and several experimental details (frame count, synchronization tolerance, statistical variability) are under-specified. The contribution of the dataset and of the ViT-Adapter-specific result is credible, but the broader claim needs substantial qualification.

major comments (4)
  1. [Section IV.C, Table II; Section V.A] The claim that the multimodal approach 'consistently surpasses unimodal methods' is contradicted by the paper's own reported results. For lane dividers with U-Net, multimodal mIoU is 0.102 versus 0.216 for image-only and 0.265 for point-cloud-only. For PidNet, multimodal lane-divider mIoU is 0.356 versus 0.532 for image-only and 0.407 for point-cloud-only; averaged over the three elements, PidNet multimodal (0.371) is below PidNet image-only (0.401). Only the ViT-Adapter row is uniformly better across all elements. The abstract's 4% and 18% gains appear to correspond to the ViT-Adapter row (image average 0.431, point-cloud average 0.293, multimodal average 0.474), not to an across-model statement. The paper must either restrict the central claim to the ViT-Adapter configuration or present per-model, per-element results with an explicit statement that fusion can degrade lane-divider segmentation for CNN backbones.
  2. [Section IV.C, Table II; Section IV.B, Table IV] The number of frames used to produce the mIoU results in Table II is never stated. This is load-bearing because Table IV shows that mIoU varies from 0.268 with 1 frame to 0.498 with 100 frames. Without specifying the frame count for each modality and model, the comparison is not reproducible, and the reported multimodal advantage could be an artifact of using a different number of fused frames than the unimodal baselines. Please state the frame count used for every reported result and ensure it is identical across compared modalities.
  3. [Section II and Section IV.A] The temporal synchronization tolerance is described inconsistently: Section II states a maximum tolerance of 50 ms based on nearest-neighbor timestamp matching, while Section IV.A states that time alignment error is controlled within 0.02 seconds (20 ms). The experiments should state which tolerance was actually used and how pairs with misalignment between 20 ms and 50 ms were handled. If the stricter bound was used, the description in Section II should be corrected; otherwise, the experimental section should explain the relaxed criterion.
  4. [Section IV.C] No error bars, confidence intervals, or repeated-run statistics are reported. With only seven test intersections and a single evaluation per configuration, differences of 0.02-0.04 in mIoU between ViT-Adapter multimodal and unimodal results may be within run-to-run or intersection-to-intersection variability. Reporting variance across training seeds or a per-intersection breakdown would strengthen the claim that the fusion improvement is reliable rather than incidental.
minor comments (7)
  1. [Section III.A, Eq. (2)] In Equation (2), the expression for y_i uses Xmin in the numerator and Ysize in the denominator; it should presumably read (y_p - Ymin)/Ysize. Please correct this typo.
  2. [Table I caption] The caption lists 'mIoU 1.000 0.431 0.293 0.474' but the table appears to show qualitative map overlays. Clarify what the numbers correspond to and how the table relates to the quantitative results in Table II.
  3. [Section V.A] The text reports 'an average improvement of about 11% in mIoU for multimodal compared to unimodal recognition,' but the abstract and conclusion report 4% and 18%. Please clarify how the 11% figure is computed and reconcile it with the other reported improvements.
  4. [Section IV.A] The paper should state whether the training data from NavInfo Company follow the same annotation protocol as RS-seq and whether any RS-seq data were used for validation or hyperparameter selection, given the statement that RS-seq was 'entirely allocated to the test set.' This is relevant for assessing possible domain shift and whether the reported numbers are optimistic.
  5. [Eq. (1)] The leading symbol 'ZC' in Equation (1) is introduced without definition; if it is the homogeneous scale factor, it should be defined in the notation list.
  6. [Figure 3] Figure 3 lacks labeled axes and units; adding explicit distance and mIoU axes would make the distance-performance trend easier to verify against Table V.
  7. [Throughout] The dataset name is written inconsistently as 'V2X-seq' and 'V2X-Seq'; please standardize the spelling.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports an empirical comparison on a held-out test set rather than a derivation from fitted inputs.

full rationale

The paper's central claims are empirical evaluations of a camera-LiDAR fusion pipeline on the RS-seq dataset, which the authors constructed and used only for testing while training on external NavInfo data. No parameter is fitted to the test set and then relabeled as a prediction; no prediction is derived from a quantity that is defined in terms of the predicted output; no uniqueness theorem or prior result by the same authors is invoked to force a choice; and the benchmark methods cited are external baselines whose use does not smuggle in the fusion result. The only nearby concerns are internal consistency issues in the reported gains (e.g., Table II shows U-Net multimodal lane-divider mIoU of 0.102 versus 0.216 image-only, which contradicts the phrase 'consistently surpasses'), but these are correctness or presentation issues, not circularity. Because the evaluation is externally anchored to a held-out dataset and no load-bearing step reduces by construction to its own inputs, the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No physically invented entities. The central claim is an empirical evaluation; the main 'free parameters' are hand-chosen processing thresholds and the unspecified frame count. The evaluation assumes accurate calibration and synchronization, and transferability of models trained on proprietary NavInfo data.

free parameters (5)
  • Grid edge length for LiDAR intensity image = 0.01 m
    Set manually in Sec. IV-A; directly controls resolution of the point cloud intensity image used for segmentation.
  • Neighbour clustering threshold = 0.5 m
    Nearest-neighbor clustering threshold in Sec. IV-A; affects vectorization of point cloud elements.
  • IoU threshold for true positive = 0.1
    Used in AP computation (Sec. IV-B3); chosen without sensitivity analysis.
  • Chamfer distance threshold = 1.0 m
    Used in AP computation; chosen without sensitivity analysis.
  • Number of frames fused = 50 (default)
    Table IV shows mIoU improves with frame count; the frame count used for Table II results is not stated.
assumptions (4)
  • standard math Pinhole camera projection (Eq. 1) and RANSAC plane fitting for ground extraction
    Standard geometric models assumed accurate after manual calibration.
  • domain assumption Temporal synchronization between camera and LiDAR, stated as 50 ms tolerance in Sec. II and 0.02 s in Sec. IV-A
    The paper gives two different sync tolerances; the fusion and evaluation depend on correct time alignment.
  • domain assumption NavInfo training data (60,000 images, 20,000 LiDAR frames) are representative of roadside intersection scenes
    Models are trained exclusively on proprietary vehicle-based NavInfo data; the RS-seq test set is from roadside installations.
  • ad hoc to paper Ground-truth HD maps use a global coordinate system with coordinates offset for data security
    Sec. II; the offset is not quantified and may affect vector-map evaluation when combined with released calibration.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal HD Mapping for Intersections by Intelligent Roadside Units." pith.science (2026). https://pith.science/paper/XLBG3MWF

@misc{pith2026250708903,
  author       = {Pith},
  title        = {Pith review of: Multimodal HD Mapping for Intersections by Intelligent Roadside Units},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XLBG3MWF}},
  note         = {Machine review of arXiv:2507.08903}
}
read the original abstract

High-definition (HD) semantic mapping of complex intersections poses significant challenges for traditional vehicle-based approaches due to occlusions and limited perspectives. This paper introduces a novel camera-LiDAR fusion framework that leverages elevated intelligent roadside units (IRUs). Additionally, we present RS-seq, a comprehensive dataset developed through the systematic enhancement and annotation of the V2X-Seq dataset. RS-seq includes precisely labelled camera imagery and LiDAR point clouds collected from roadside installations, along with vectorized maps for seven intersections annotated with detailed features such as lane dividers, pedestrian crossings, and stop lines. This dataset facilitates the systematic investigation of cross-modal complementarity for HD map generation using IRU data. The proposed fusion framework employs a two-stage process that integrates modality-specific feature extraction and cross-modal semantic integration, capitalizing on camera high-resolution texture and precise geometric data from LiDAR. Quantitative evaluations using the RS-seq dataset demonstrate that our multimodal approach consistently surpasses unimodal methods. Specifically, compared to unimodal baselines evaluated on the RS-seq dataset, the multimodal approach improves the mean Intersection-over-Union (mIoU) for semantic segmentation by 4\% over the image-only results and 18\% over the point cloud-only results. This study establishes a baseline methodology for IRU-based HD semantic mapping and provides a valuable dataset for future research in infrastructure-assisted autonomous driving systems.

Figures

Figures reproduced from arXiv: 2507.08903 by the authors.

Figure 1
Figure 1. Multimodal data acquisition and semantic HD map generation framework utilizing intelligent roadside units: integrating camera imagery and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The systematic framework for semantic HD map generation through intelligent roadside unit data processing: a pipeline from multimodal data [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The impact of varying distances to sensors on map quality. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages

  1. [1]

    Hdmapnet: An online hd map construction and evaluation framework,

    Q. Li, Y . Wang, Y . Wang, and H. Zhao, “Hdmapnet: An online hd map construction and evaluation framework,” in 2022 International Conference on Robotics and Automation (ICRA) . IEEE, 2022, pp. 4628–4634

    work page 2022

  2. [2]

    Maptrv2: An end-to-end framework for online vectorized hd map construction,

    B. Liao, S. Chen, Y . Zhang, B. Jiang, Q. Zhang, W. Liu, C. Huang, and X. Wang, “Maptrv2: An end-to-end framework for online vectorized hd map construction,” arXiv preprint arXiv:2308.05736 , 2023

    arXiv 2023

  3. [3]

    Mgmap: Mask-guided learning for online vectorized hd map construction,

    X. Liu, S. Wang, W. Li, R. Yang, J. Chen, and J. Zhu, “Mgmap: Mask-guided learning for online vectorized hd map construction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 14 812–14 821

    2024

  4. [4]

    V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting,

    H. Yu, W. Yang, H. Ruan, Z. Yang, Y . Tang, X. Gao, X. Hao, Y . Shi, Y . Pan, N. Sun, et al. , “V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 5486–5495

    work page 2023

  5. [5]

    Design and implementation of edge-fog-cloud system through hd map generation from lidar data of autonomous vehicles,

    J. Lee, K. Lee, A. Yoo, and C. Moon, “Design and implementation of edge-fog-cloud system through hd map generation from lidar data of autonomous vehicles,” Electronics, vol. 9, no. 12, p. 2084, 2020

    work page 2020

  6. [6]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , June 2019

    work page 2019

  7. [7]

    Bevfusion: A simple and robust lidar-camera fusion framework,

    T. Liang, H. Xie, K. Yu, Z. Xia, Z. Lin, Y . Wang, T. Tang, B. Wang, and Z. Tang, “Bevfusion: A simple and robust lidar-camera fusion framework,” Advances in Neural Information Processing Systems , vol. 35, pp. 10 421–10 434, 2022

    work page 2022

  8. [8]

    Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,

    J. Philion and S. Fidler, “Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16 . Springer, 2020, pp. 194–210

    work page 2020

Show all 22 references
  1. [9]

    Second: Sparsely embedded convolutional detection,

    Y . Yan, Y . Mao, and B. Li, “Second: Sparsely embedded convolutional detection,” Sensors, vol. 18, no. 10, p. 3337, 2018

    2018

  2. [10]

    Superfusion: Multilevel lidar-camera fusion for long- range hd map generation,

    H. Dong, W. Gu, X. Zhang, J. Xu, R. Ai, H. Lu, J. Kannala, and X. Chen, “Superfusion: Multilevel lidar-camera fusion for long- range hd map generation,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) , 2024, pp. 9056–9062

    2024

  3. [11]

    V2x-vit: Vehicle-to-everything cooperative perception with vision transformer,

    R. Xu, H. Xiang, Z. Tu, X. Xia, M.-H. Yang, and J. Ma, “V2x-vit: Vehicle-to-everything cooperative perception with vision transformer,” in European conference on computer vision . Springer, 2022, pp. 107– 124

    2022

  4. [12]

    Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task,

    X. Ye, M. Shu, H. Li, Y . Shi, Y . Li, G. Wang, X. Tan, and E. Ding, “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 (CVPR), June 2022,...

    2022

  5. [13]

    Dair-v2x: A large-scale dataset for vehicle- infrastructure cooperative 3d object detection,

    H. Yu, Y . Luo, M. Shu, Y . Huo, Z. Yang, Y . Shi, Z. Guo, H. Li, X. Hu, J. Yuan, 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 , 2022...

    2022

  6. [14]

    Tumtraf v2x cooperative perception dataset,

    W. Zimmer, G. A. Wardana, S. Sritharan, X. Zhou, R. Song, and A. C. Knoll, “Tumtraf v2x cooperative perception dataset,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. IEEE, 2024

    2024

  7. [15]

    Vi-map: Infrastructure-assisted real-time hd mapping for autonomous driving,

    Y . He, C. Bian, J. Xia, S. Shi, Z. Yan, Q. Song, and G. Xing, “Vi-map: Infrastructure-assisted real-time hd mapping for autonomous driving,” in Proceedings of the 29th Annual International Conference on Mobile Computing and Networking , 2023, pp. 1–15

    2023

  8. [16]

    Carla: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “Carla: An open urban driving simulator,” in Conference on robot learning . PMLR, 2017, pp. 1–16

    2017

  9. [17]

    Denoising of a multi- station point cloud and 3d modeling accuracy for substation equipment based on statistical outlier removal,

    J. Guo, W. Feng, T. Hao, P. Wang, and H. Mao, “Denoising of a multi- station point cloud and 3d modeling accuracy for substation equipment based on statistical outlier removal,” in 2020 IEEE 4th Conference on Energy Internet and Energy System Integration (EI2) , 2020

    2020

  10. [18]

    Three-dimensional alpha shapes,

    H. Edelsbrunner and E. P. Mücke, “Three-dimensional alpha shapes,” ACM Transactions on Graphics , vol. 13, no. 1, 1994

    1994

  11. [19]

    Least-squares fitting of a straight line,

    D. York, “Least-squares fitting of a straight line,” Canadian Journal of Physics , vol. 44, no. 5, pp. 1079–1086, 1966

    1966

  12. [20]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, pro- ceedings, part III 18 ...

    2015

  13. [21]

    Pidnet: A real-time semantic segmentation network inspired by pid controllers,

    J. Xu, Z. Xiong, and S. P. Bhattacharyya, “Pidnet: A real-time semantic segmentation network inspired by pid controllers,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 19 529–19 539

    2023

  14. [22]

    Vision transformer adapter for dense predictions,

    Z. Chen, Y . Duan, W. Wang, J. He, T. Lu, J. Dai, and Y . Qiao, “Vision transformer adapter for dense predictions,” in The Eleventh International Conference on Learning Representations

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.