Pith. sign in

REVIEW 4 major objections 5 minor 55 references

Domain Adaptation-Based Crossmodal Knowledge Distillation for 3D Semantic Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that 3D LiDAR semantic segmentation can be trained from 2D image knowledge alone via crossmodal distillation, reaching 63.7 mIoU on nuScenes without any 3D labels in the zero-shot setting.

desk verdict Useful method for low-annotation 3D segmentation, but the zero-shot result is undercut by an unexplained inconsistency in the annotation-efficiency table. read the letter →

arxiv 2509.00379 v1 pith:BPVVHKYB submitted 2025-08-30 cs.CV cs.RO

Jialiang Kang , Jiawen Wang , Dingsheng Luo This is my paper
classification cs.CVcs.RO
keywords 3DsemanticsegmentationLiDARpointcloudscrossmodalknowledgedistillationdomainadaptationself-calibratedconvolutionzero-shotfew-shotautonomousdriving
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 remove the need for expensive 3D point-cloud annotations in LiDAR semantic segmentation by transferring knowledge from 2D image networks, which are cheaper to train and have abundant data. It proposes two distillation methods: UDAKD, for unlabeled images, aligns 3D point features with 2D image features using contrastive learning over superpixels; FSKD, for labeled images, additionally aligns semantic soft labels through a shared classifier. The key move is a domain adaptation module built from 3D self-calibrated convolutions, which maps LiDAR features into a pseudo-2D feature space while keeping modality-general information and suppressing modality-specific details. On nuScenes, FSKD reports 63.7 mIoU for zero-shot 3D segmentation (no 3D labels) and UDAKD reaches 39.3 mIoU after 1% fine-tuning, both ahead of previous state-of-the-art methods. If correct, the result matters because it points to a practical route where image data alone—or with very few point-cloud labels—can teach a 3D network to segment.

What carries the argument

The load-bearing component is the domain adaptation module mλ built from multi-layered 3D self-calibrated convolutions. It transforms raw 3D features G into pseudo-2D features G′ that live in the same space as the 2D teacher's features F, so the student can be supervised at the feature level; the module is what filters out modality-specific details while keeping modality-general semantics. Two other mechanisms carry the transfer: (1) superpixel/superpoint pooling in UDAKD, which creates robust correspondences between 2D pixels and sparse 3D points before computing the InfoNCE loss; and (2) a shared MLP classifier in FSKD, which, under the neural-collapse assumption, turns aligned 3D features

What would settle it

Take a FSKD-trained 3D feature extractor and measure neural collapse directly: compute the within-class variance relative to the distance between class means on the nuScenes validation set. If features from the same class are not tightly clustered, then the shared classifier is not actually transferring semantics; a second, cleaner test is training a new 2D classifier on a disjoint set of fine-grained classes and checking whether the 3D network can label them without any 3D fine-tuning—if per-class mIoU collapses, the zero-shot claim does not generalize.

Watch

Extended reading notes

Core claim

The central discovery is that a 3D LiDAR segmentation network can learn semantic structure from a 2D image teacher without any 3D supervision, provided the two feature spaces are actively aligned. The paper introduces UDAKD and FSKD. UDAKD uses a frozen self-supervised 2D model and distills feature-level knowledge through an InfoNCE loss on matched superpixel-superpoint regions. FSKD uses a fully supervised 2D segmentation network, and distills both features (MSE) and semantic soft labels (KL) into the 3D student while sharing one MLP classifier across modalities. Under the neural-collapse assumption—features of each class cluster at their mean—the shared classifier lets the 3D network inher

Load-bearing premise

The claim rests on the assumption that both the 2D and 3D feature extractors undergo complete neural collapse—that is, all features of a class converge to a single point—so one shared classifier can correctly label features from either modality; the paper itself admits perfect neural collapse is impractical and relies on the domain adaptation module to compensate for the gap.

Editorial extensions

If this is right

  • If FSKD's zero-shot number holds, a 3D segmentation model can be fielded for a new scene type using only a labeled 2D dataset and synchronized camera-LiDAR capture, skipping point-cloud labeling entirely.
  • UDAKD supplies a pretraining recipe that improves few-shot fine-tuning: at 1% labels it adds 1.0 mIoU over SLidR on nuScenes and 0.5 on SemanticKITTI, and the gain persists even at 100% labels (+0.4 over SLidR).
  • The shared classifier makes the 3D network able to reuse the 2D network's updated classifier, so new fine-grained classes such as debris, bicycle racks, and construction workers can be segmented by the 3D network without retraining or 3D labels.
  • Removing the DA module or the semantic distillation sharply reduces zero-shot FSKD (by 15.2 and 15.7 mIoU, respectively), so the alignment module, not the teacher alone, carries the transfer.

Reading between the lines

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

  • This suggests the same 3D self-calibrated convolution module could be lifted to other sparse 3D tasks such as object detection or panoptic segmentation, and to other sensor pairs such as radar-to-LiDAR, since the module only assumes a known correspondence and a shared feature space.
  • The dependence on neural collapse implies the zero-shot gain may shrink on long-tailed or highly variable classes; a stress test would be running FSKD with a 2D teacher trained on a heavily imbalanced label set and measuring per-class mIoU.
  • The paper leaves the calibration problem open; one extension is to learn the 2D-3D correspondence instead of assuming perfect extrinsic calibration, which would widen the method to platforms with loose sensor synchronization.
  • FSKD's 63.7 mIoU is reported on a shared 11-class taxonomy; testing on the full 16-class nuScenes map or on completely novel vocabularies would tell whether the shared classifier's transfer is limited by class granularity.
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 / 5 minor

Summary. The paper proposes two crossmodal knowledge distillation methods, UDAKD and FSKD, for 3D semantic segmentation from LiDAR point clouds. Both exploit synchronized 2D images and 3D LiDAR data to transfer knowledge from a 2D image model to a 3D network, using a domain adaptation module based on 3D self-calibrated convolutions. UDAKD is designed for unlabeled images and uses a frozen self-supervised 2D backbone; FSKD assumes a 2D segmentation network trained on labeled images. The methods are evaluated for few-shot and zero-shot 3D semantic segmentation on nuScenes and SemanticKITTI, reporting state-of-the-art mIoU in several settings.

Significance. If the reported results are reliable, the paper makes a useful contribution to reducing 3D annotation costs by demonstrating that crossmodal distillation, augmented with a 3D domain adaptation module, can transfer semantic knowledge from 2D images to LiDAR point clouds. The inclusion of a shared 2D/3D classifier for zero-shot segmentation is an interesting extension of prior work (e.g., SLidR, PPKT, CMKD). The paper provides code and extensive ablations, which are strengths. However, the central comparative claims depend on experimental protocols that raise concerns about fairness and internal consistency; these must be resolved before the results can be accepted.

major comments (4)
  1. [§IV-D-a, Table III] Table III reports a non-monotonic trend for FSKD: mIoU is 63.7 at 0% annotations but drops to 62.0 at 1%, while SLidR rises from 4.1 to 32.9. This contradicts the text's statement that 'as annotation proportions increase, both methods improve.' The 1% drop is unexplained and undermines the reliability of the zero-shot headline number. Please clarify the evaluation protocol for each row: is the 0% row obtained with the shared 2D/3D classifier (as in Table II) while the 1% row uses a newly trained classifier on 1% labels? If so, the rows are not directly comparable and the claim of annotation efficiency needs to be re-evaluated. Also explain how SLidR is evaluated at 0% (no 3D labels), since training any classifier requires some supervision.
  2. [§IV-B-c, Table I] The comparison in Table I is not apples-to-apples. UDAKD and FSKD use 2D teachers that are either self-supervised (UDAKD) or fully supervised on nuImages (FSKD), while the baselines I2P-MAE, PPKT, and SLidR use self-supervised backbones as teachers. The FSKD gain over SLidR (+6.4 mIoU at 1%) may largely stem from the stronger supervised teacher, not from the proposed DA module or distillation losses. The paper's own acknowledgment that 'this approach might seem biased' does not address the confound. Please provide a controlled comparison: e.g., FSKD with a self-supervised 2D teacher (same protocol as SLidR), or SLidR/PPKT with a supervised teacher, so that the effect of the proposed components is isolated.
  3. [§IV (all tables)] No error bars or multiple-seed results are reported. Given the non-monotonicity in Table III and the small differences in some comparisons (e.g., UDAKD vs. SLidR in Table I), single runs are insufficient to establish the claimed improvements. Please run each experiment with at least 3 random seeds and report mean ± std, or at least state the variance if a single seed was used. This is particularly important for the few-shot and zero-shot settings, where initialization can have a large effect.
  4. [§III-C] The zero-shot transfer mechanism relies on the assumption of 'complete neural collapse for both 2D and 3D extractors,' which the paper acknowledges is impractical. This is a correctness risk: if the features are not sufficiently collapsed, the shared classifier can produce misaligned predictions. The paper introduces a DA module to mitigate this, but no direct evidence is provided that the assumption holds to the required degree. Please add a quantitative test of neural collapse (e.g., the within-class variance to between-class variance ratio) for the 2D and 3D features used in the zero-shot evaluation, or a sensitivity analysis showing that the result is robust to partial collapse. This would make the method's reliance on the assumption explicit and credible.
minor comments (5)
  1. [§III-C] The description of when the shared classifier is used is ambiguous. In FSKD, is the classifier always shared between 2D and 3D, or only in the zero-shot setting? The text says 'For zero-shot domain adaptation' the parameters are shared, but the method overview in Fig. 1 seems to show a shared classifier for FSKD in general. Please clarify in the text.
  2. [§IV-B-a] The training details state that 'damping' is set to 0.1. For SGD, this is unconventional; if this refers to the learning rate scheduler's decay or a momentum damping term, please specify the exact optimizer and parameters used.
  3. [§II-B] The related work section introduces CMKD [11] as a 2D-domain DA module, but the paper then claims the DA module's applicability in 3D is 'unexplored.' This is fine, but the connection between the proposed 3D self-calibrated convolution and the 2D version in [10] should be stated more precisely in the method section, including how the convolution is adapted for sparse voxel grids.
  4. [§IV-D-b, Table IV] In Table IV, the notation 'FSKD*' and 'SegNet*' is not defined in the table caption or text. Please define the asterisk notation: does it indicate the retrained classifier setting? Also, the row labels for classes are inconsistent with the column headers; please align them.
  5. [§IV-C] When describing the 2DPASS baseline, the text mentions 'a fixed classifier trained on 2D images.' Please specify which classifier this is and whether it is the shared classifier used in FSKD. Without this, the comparison in Table II is difficult to interpret.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical, benchmarked against external methods, and contain no fitted-input-as-prediction or self-citation load-bearing steps.

full rationale

The paper does not claim a first-principles derivation. Its central claims are empirical: UDAKD and FSKD are training procedures that transfer 2D knowledge to 3D networks via feature and semantic distillation, evaluated on nuScenes/SemanticKITTI against external baselines. No parameter is fitted to the target evaluation and then reported as a prediction; the 3D network is never trained with 3D labels in the zero-shot setup, and the few-shot results are after fine-tuning on the same labels available to all baselines. The shared classifier in FSKD is taken from the 2D network, not derived from or fitted to the 3D validation data; its justification is an externally cited neural-collapse phenomenon ([39]) whose failure the paper explicitly acknowledges. Citations to prior work (SLidR, PPKT, I2P-MAE, CMKD, etc.) are external and non-overlapping with the author list; there is no self-citation chain or imported uniqueness theorem. The DA module uses self-calibrated convolution from [10], an external component, and its contribution is tested by ablations rather than assumed. The Table III non-monotonicity (FSKD 63.7 at 0% vs 62.0 at 1%) is a potential evaluation-consistency or robustness issue, but it is not a circularity: it does not show that the reported zero-shot number is equivalent by construction to a fitted input. Therefore the circularity score is 0.

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

The method is empirical; its central assumptions are the shared-classifier neural-collapse claim and known sensor calibration. All losses and hyperparameters are standard for the knowledge distillation literature. No new physical entities are introduced.

free parameters (4)
  • FSKD loss weighting factors a and b = 10:1 (feature:semantic)
    Set in Section IV-B; tuned on a 100-scene validation split. Affects the balance between feature and semantic distillation.
  • InfoNCE temperature tau = 0.07
    Standard value for contrastive losses, chosen in Section IV-B.
  • Maximum SLIC superpixels per image = 150
    Set in Section IV-B; controls the granularity of UDAKD feature pooling and distillation.
  • DA module depth = 3-layer self-calibrated convolution
    Design choice in Section IV-B; not varied in depth ablation, only removed entirely to test its importance.
assumptions (4)
  • domain assumption Complete neural collapse of both 2D and 3D features
    Section III-C states 'we assume complete neural collapse for both 2D and 3D extractors' to justify a shared classifier for zero-shot transfer.
  • domain assumption Known extrinsic calibration between camera and LiDAR
    Section III-E and the conclusion rely on 'reliable, one-time extrinsic parameter calibration' to establish 2D-3D correspondence.
  • domain assumption Pretrained 2D feature extractor availability
    UDAKD uses an ImageNet-pretrained ResNet-50 with MoCo v2; FSKD trains SegNet on nuImages. The method assumes such a teacher exists.
  • domain assumption nuImages labels are a valid source for the 2D teacher
    FSKD relies on nuImages 11-class annotations to train the 2D teacher; no 3D labels are used in distillation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Domain Adaptation-Based Crossmodal Knowledge Distillation for 3D Semantic Segmentation." pith.science (2026). https://pith.science/paper/BPVVHKYB

@misc{pith2026250900379,
  author       = {Pith},
  title        = {Pith review of: Domain Adaptation-Based Crossmodal Knowledge Distillation for 3D Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BPVVHKYB}},
  note         = {Machine review of arXiv:2509.00379}
}
read the original abstract

Semantic segmentation of 3D LiDAR data plays a pivotal role in autonomous driving. Traditional approaches rely on extensive annotated data for point cloud analysis, incurring high costs and time investments. In contrast, realworld image datasets offer abundant availability and substantial scale. To mitigate the burden of annotating 3D LiDAR point clouds, we propose two crossmodal knowledge distillation methods: Unsupervised Domain Adaptation Knowledge Distillation (UDAKD) and Feature and Semantic-based Knowledge Distillation (FSKD). Leveraging readily available spatio-temporally synchronized data from cameras and LiDARs in autonomous driving scenarios, we directly apply a pretrained 2D image model to unlabeled 2D data. Through crossmodal knowledge distillation with known 2D-3D correspondence, we actively align the output of the 3D network with the corresponding points of the 2D network, thereby obviating the necessity for 3D annotations. Our focus is on preserving modality-general information while filtering out modality-specific details during crossmodal distillation. To achieve this, we deploy self-calibrated convolution on 3D point clouds as the foundation of our domain adaptation module. Rigorous experimentation validates the effectiveness of our proposed methods, consistently surpassing the performance of state-of-the-art approaches in the field.

Figures

Figures reproduced from arXiv: 2509.00379 by the authors.

Figure 1
Figure 1. Illustration of UDAKD and FSKD. Features are generated from 3D point clouds and 2D images through their respective networks. The 3D features [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of 3D feature extractor outputs pretrained with [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Visualization of 3D features with color-coded [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

55 extracted references · 42 canonical work pages

  1. [1]

    nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 11 621–11 631

    2020

  2. [2]

    Vision meets robotics: The kitti dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The kitti dataset,” The International Journal of Robotics Research , vol. 32, no. 11, pp. 1231–1237, 2013

    2013

  3. [3]

    One thousand and one hours: Self-driving motion prediction dataset,

    J. Houston, G. Zuidhof, L. Bergamini, Y . Ye, L. Chen, A. Jain, S. Omari, V . Iglovikov, and P. Ondruska, “One thousand and one hours: Self-driving motion prediction dataset,” in Conference on Robot Learning. PMLR, 2021, pp. 409–418

    work page 2021

  4. [4]

    Scalability in perception for autonomous driving: Waymo open dataset,

    P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V . Patnaik, P. Tsui, J. Guo, Y . Zhou, Y . Chai, B. Caine et al. , “Scalability in perception for autonomous driving: Waymo open dataset,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 2446–2454

    2020

  5. [5]

    Learning 3d semantic segmentation with only 2d image supervision,

    K. Genova, X. Yin, A. Kundu, C. Pantofaru, F. Cole, A. Sud, B. Brewington, B. Shucker, and T. Funkhouser, “Learning 3d semantic segmentation with only 2d image supervision,” in 2021 International Conference on 3D Vision (3DV) . IEEE, 2021, pp. 361–372

    work page 2021

  6. [6]

    Image-to-lidar self-supervised distillation for autonomous driving data,

    C. Sautier, G. Puy, S. Gidaris, A. Boulch, A. Bursuc, and R. Marlet, “Image-to-lidar self-supervised distillation for autonomous driving data,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 9891–9901

    work page 2022

  7. [7]

    Complete & label: A domain adaptation approach to semantic segmentation of lidar point clouds,

    L. Yi, B. Gong, and T. Funkhouser, “Complete & label: A domain adaptation approach to semantic segmentation of lidar point clouds,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 15 363–15 373

    work page 2021

  8. [8]

    Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining,

    Y .-C. Liu, Y .-K. Huang, H.-Y . Chiang, H.-T. Su, Z.-Y . Liu, C.-T. Chen, C.-Y . Tseng, and W. H. Hsu, “Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining,” arXiv preprint arXiv:2104.04687, 2021

    arXiv 2021

Show all 55 references
  1. [9]

    The modality focusing hy- pothesis: Towards understanding crossmodal knowledge distillation,

    Z. Xue, Z. Gao, S. Ren, and H. Zhao, “The modality focusing hy- pothesis: Towards understanding crossmodal knowledge distillation,” arXiv preprint arXiv:2206.06487 , 2022

    2022 arXiv

  2. [10]

    Improving convolutional networks with self-calibrated convolutions,

    J.-J. Liu, Q. Hou, M.-M. Cheng, C. Wang, and J. Feng, “Improving convolutional networks with self-calibrated convolutions,” in Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 10 096–10 105

    2020

  3. [11]

    Cross-modality knowledge distillation network for monocular 3d object detection,

    Y . Hong, H. Dai, and Y . Ding, “Cross-modality knowledge distillation network for monocular 3d object detection,” in Computer Vision– ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23– 27, 2022, Proceedings, Part X . Springer, 2022, pp. 87–104

    2022

  4. [12]

    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

  5. [13]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

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

    2017

  6. [14]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,” Advances in neural information processing systems , vol. 30, 2017

    2017

  7. [15]

    Deep parametric continuous convolutional neural networks,

    S. Wang, S. Suo, W.-C. Ma, A. Pokrovsky, and R. Urtasun, “Deep parametric continuous convolutional neural networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 2589–2597

    2018

  8. [16]

    Kpconv: Flexible and deformable convolution for point clouds,

    H. Thomas, C. R. Qi, J.-E. Deschaud, B. Marcotegui, F. Goulette, and L. J. Guibas, “Kpconv: Flexible and deformable convolution for point clouds,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6411–6420

    2019

  9. [17]

    Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud,

    B. Wu, A. Wan, X. Yue, and K. Keutzer, “Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud,” in 2018 IEEE international conference on robotics and automation (ICRA) . IEEE, 2018, pp. 1887–1893

    2018

  10. [18]

    Squeezesegv2: Im- proved model structure and unsupervised domain adaptation for road- object segmentation from a lidar point cloud,

    B. Wu, X. Zhou, S. Zhao, X. Yue, and K. Keutzer, “Squeezesegv2: Im- proved model structure and unsupervised domain adaptation for road- object segmentation from a lidar point cloud,” in 2019 International Conference on Robotics and Automation (ICRA) . IEEE, 2019, pp. 4376–4382

    2019

  11. [19]

    Squeezesegv3: Spatially-adaptive convolution for ef- ficient point-cloud segmentation,

    C. Xu, B. Wu, Z. Wang, W. Zhan, P. Vajda, K. Keutzer, and M. Tomizuka, “Squeezesegv3: Spatially-adaptive convolution for ef- ficient point-cloud segmentation,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part XXVIII 1...

    2020

  12. [20]

    Rangenet++: Fast and accurate lidar semantic segmentation,

    A. Milioto, I. Vizzo, J. Behley, and C. Stachniss, “Rangenet++: Fast and accurate lidar semantic segmentation,” in 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS) . IEEE, 2019, pp. 4213–4220

    2019

  13. [21]

    Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds,

    T. Cortinhal, G. Tzelepis, and E. Erdal Aksoy, “Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds,” in Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5–7, 2020, Proceedings, Part II

    2020

  14. [22]

    Springer, 2020, pp. 207–222

    2020

  15. [23]

    Polarnet: An improved grid representation for online lidar point clouds semantic segmentation,

    Y . Zhang, Z. Zhou, P. David, X. Yue, Z. Xi, B. Gong, and H. Foroosh, “Polarnet: An improved grid representation for online lidar point clouds semantic segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 9601–9610

    2020

  16. [24]

    Panoptic-polarnet: Proposal- free lidar point cloud panoptic segmentation,

    Z. Zhou, Y . Zhang, and H. Foroosh, “Panoptic-polarnet: Proposal- free lidar point cloud panoptic segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 13 194–13 203

    2021

  17. [25]

    4d spatio-temporal convnets: Minkowski convolutional neural networks,

    C. Choy, J. Gwak, and S. Savarese, “4d spatio-temporal convnets: Minkowski convolutional neural networks,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 3075–3084

    2019

  18. [26]

    Pointcon- trast: Unsupervised pre-training for 3d point cloud understanding,

    S. Xie, J. Gu, D. Guo, C. R. Qi, L. Guibas, and O. Litany, “Pointcon- trast: Unsupervised pre-training for 3d point cloud understanding,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16 . Springer, 2020, pp. 574–591

    2020

  19. [27]

    Cylindrical and asymmetrical 3d convolution networks for lidar segmentation,

    X. Zhu, H. Zhou, T. Wang, F. Hong, Y . Ma, W. Li, H. Li, and D. Lin, “Cylindrical and asymmetrical 3d convolution networks for lidar segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 9939–9948

    2021

  20. [28]

    (af)2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network,

    R. Cheng, R. Razani, E. Taghavi, E. Li, and B. Liu, “(af)2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 12 547–12 556

    2021

  21. [29]

    Search- ing efficient 3d architectures with sparse point-voxel convolution,

    H. Tang, Z. Liu, S. Zhao, Y . Lin, J. Lin, H. Wang, and S. Han, “Search- ing efficient 3d architectures with sparse point-voxel convolution,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII . Springer, 2020, pp. 685–702

    2020

  22. [30]

    Svqnet: Sparse voxel-adjacent query network for 4d spatio-temporal lidar se- mantic segmentation,

    X. Chen, S. Xu, X. Zou, T. Cao, D.-Y . Yeung, and L. Fang, “Svqnet: Sparse voxel-adjacent query network for 4d spatio-temporal lidar se- mantic segmentation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 8569–8578

    2023

  23. [31]

    Cross modal distillation for supervision transfer,

    S. Gupta, J. Hoffman, and J. Malik, “Cross modal distillation for supervision transfer,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 2827–2836

    2016

  24. [32]

    Improving unimodal object recognition with multimodal contrastive learning,

    J. Meyer, A. Eitel, T. Brox, and W. Burgard, “Improving unimodal object recognition with multimodal contrastive learning,” in 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 5656–5663

    2020

  25. [33]

    Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders,

    R. Zhang, L. Wang, Y . Qiao, P. Gao, and H. Li, “Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , 2023, pp. 21 769–21 780

    2023

  26. [34]

    Weakly supervised 3d multi-person pose estimation for large-scale scenes based on monocular camera and single lidar,

    P. Cong, Y . Xu, Y . Ren, J. Zhang, L. Xu, J. Wang, J. Yu, and Y . Ma, “Weakly supervised 3d multi-person pose estimation for large-scale scenes based on monocular camera and single lidar,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 37, 2023, pp. 461–469

    2023

  27. [35]

    xmuda: Cross-modal unsupervised domain adaptation for 3d semantic seg- mentation,

    M. Jaritz, T.-H. Vu, R. d. Charette, E. Wirbel, and P. P ´erez, “xmuda: Cross-modal unsupervised domain adaptation for 3d semantic seg- mentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 12 605–12 614

    2020

  28. [36]

    Cross- modal learning for domain adaptation in 3d semantic segmentation,

    M. Jaritz, T.-H. Vu, R. De Charette, ´E. Wirbel, and P. P ´erez, “Cross- modal learning for domain adaptation in 3d semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 2, pp. 1533–1544, 2022

    2022

  29. [37]

    Openscene: 3d scene understanding with open vocabularies,

    S. Peng, K. Genova, C. Jiang, A. Tagliasacchi, M. Pollefeys, T. Funkhouser et al. , “Openscene: 3d scene understanding with open vocabularies,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 815–824

    2023

  30. [38]

    Slic superpixels compared to state-of-the-art superpixel methods,

    R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. S ¨usstrunk, “Slic superpixels compared to state-of-the-art superpixel methods,” IEEE transactions on pattern analysis and machine intelligence , vol. 34, no. 11, pp. 2274–2282, 2012

    2012

  31. [39]

    Image2point: 3d point-cloud understanding with 2d image pretrained models,

    C. Xu, S. Yang, T. Galanti, B. Wu, X. Yue, B. Zhai, W. Zhan, P. Vajda, K. Keutzer, and M. Tomizuka, “Image2point: 3d point-cloud understanding with 2d image pretrained models,” in Computer Vision– ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23– 27, 2022, Pro...

    2022

  32. [40]

    On the role of neural collapse in transfer learning,

    T. Galanti, A. Gy ¨orgy, and M. Hutter, “On the role of neural collapse in transfer learning,” arXiv preprint arXiv:2112.15121 , 2021

    2021 arXiv

  33. [41]

    Tri-perspective view for vision-based 3d semantic occupancy prediction,

    Y . Huang, W. Zheng, Y . Zhang, J. Zhou, and J. Lu, “Tri-perspective view for vision-based 3d semantic occupancy prediction,” in Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9223–9232

    2023

  34. [42]

    Segnet: A deep convolutional encoder-decoder architecture for image segmentation,

    V . Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: A deep convolutional encoder-decoder architecture for image segmentation,” IEEE transactions on pattern analysis and machine intelligence , vol. 39, no. 12, pp. 2481–2495, 2017

    2017

  35. [43]

    Mseg3d: Multi-modal 3d semantic segmentation for autonomous driving,

    J. Li, H. Dai, H. Han, and Y . Ding, “Mseg3d: Multi-modal 3d semantic segmentation for autonomous driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 21 694–21 704

    2023

  36. [44]

    Rgb and lidar fusion based 3d semantic segmentation for autonomous driving,

    K. El Madawi, H. Rashed, A. El Sallab, O. Nasr, H. Kamel, and S. Yogamani, “Rgb and lidar fusion based 3d semantic segmentation for autonomous driving,” in 2019 IEEE Intelligent Transportation Systems Conference (ITSC) . IEEE, 2019, pp. 7–12

    2019

  37. [45]

    Sensor fusion for joint 3d object detection and semantic segmentation,

    G. P. Meyer, J. Charland, D. Hegde, A. Laddha, and C. Vallespi- Gonzalez, “Sensor fusion for joint 3d object detection and semantic segmentation,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition workshops , 2019, pp. 0–0

    2019

  38. [46]

    Pointpainting: Sequential fusion for 3d object detection,

    S. V ora, A. H. Lang, B. Helou, and O. Beijbom, “Pointpainting: Sequential fusion for 3d object detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 4604–4612

    2020

  39. [47]

    Fuseseg: Lidar point cloud segmentation fusing multi-modal data,

    G. Krispel, M. Opitz, G. Waltner, H. Possegger, and H. Bischof, “Fuseseg: Lidar point cloud segmentation fusing multi-modal data,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2020, pp. 1874–1883

    2020

  40. [48]

    Multiscale deep learning network with self-calibrated convolution for hyperspectral and lidar data collaborative classification,

    Z. Xue, X. Yu, X. Tan, B. Liu, A. Yu, and X. Wei, “Multiscale deep learning network with self-calibrated convolution for hyperspectral and lidar data collaborative classification,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–16, 2021

    2021

  41. [49]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,”arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  42. [50]

    Semantickitti: A dataset for semantic scene un- derstanding of lidar sequences,

    J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stach- niss, and J. Gall, “Semantickitti: A dataset for semantic scene un- derstanding of lidar sequences,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 9297–9307

    2019

  43. [51]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition . Ieee, 2009, pp. 248–255

    2009

  44. [52]

    Improved baselines with momentum contrastive learning,

    X. Chen, H. Fan, R. Girshick, and K. He, “Improved baselines with momentum contrastive learning,” arXiv preprint arXiv:2003.04297 , 2020

    2003 arXiv

  45. [53]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hinton et al., “Learning multiple layers of features from tiny images,” Master’s thesis, University of Toronto , 2009

    2009

  46. [54]

    2dpass: 2d priors assisted semantic segmentation on lidar point clouds,

    X. Yan, J. Gao, C. Zheng, C. Zheng, R. Zhang, S. Cui, and Z. Li, “2dpass: 2d priors assisted semantic segmentation on lidar point clouds,” in Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXVIII . Springer, 2022, ...

    2022

  47. [55]

    Dense 3d semantic mapping of indoor scenes from rgb-d images,

    A. Hermans, G. Floros, and B. Leibe, “Dense 3d semantic mapping of indoor scenes from rgb-d images,” in 2014 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2014, pp. 2631–2638

    2014

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.