Pith. sign in

REVIEW 3 major objections 5 minor 36 references

TOLiD: Bridging the Architecture Gap in Vision Foundation Model to LiDAR Pretraining via Token Lifting for Distillation

T0 review · 3 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Matching LiDAR features to a same-family Vision Transformer student, not a frozen image teacher alone, yields more transferable 3D representations for segmentation.

desk verdict Solid systems paper that keeps a student ViT in the distillation loop and shows real frozen-backbone and cross-sensor gains; the architecture-gap story is plausible but not cleanly isolated from extra capacity. read the letter →

arxiv 2607.10762 v1 pith:UKMUKT2I submitted 2026-07-12 cs.CV cs.LGcs.RO

Sutharsan Mahendran , Darshana Priyasad , Kaushik Roy , Tharindu Fernando , Sridha Sridharan , Clinton Fookes , Peyman Moghadam This is my paper
classification cs.CVcs.LGcs.RO
keywords LiDARrepresentationlearningcross-modaldistillationVisionFoundationModelsTransformerfrustumattentionself-supervisedpretraining3Dsemanticsegmentationcross-sensordomainadaptation
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

Prior methods that distill Vision Foundation Models into LiDAR networks force a sparse 3D encoder to copy dense ViT embeddings, so the student must close both a sensor gap and an architecture gap at once. TOLiD instead attaches a LiDAR backbone to a student Vision Transformer initialized from the frozen teacher and trains only over compatible patch tokens. Point features inside each image-patch frustum become tokens via mean pooling and content-aware frustum attention; distillation uses visibility masks so empty image regions do not train the student. For pure LiDAR use, tokens are lifted back to points with masked bilinear sampling. Across five camera–LiDAR datasets and four cross-sensor pairs, the resulting frozen features plus a light head beat prior distillation baselines on linear probing, label-efficient finetuning, domain adaptation, and sensor-corruption robustness, and the gap widens as pretraining data grows.

What carries the argument

Token lifting for distillation: Frustum Pooling and Frustum Attention map variable point sets inside each image-patch frustum into ViT-compatible tokens; visibility-masked token distillation aligns them to the teacher; masked bilinear sampling lifts tokens back to per-point features without empty-patch contamination.

What would settle it

Pretrain under the same multi-dataset recipe but with deliberately noisy or time-misaligned extrinsics (or without cameras at all) and check whether the linear-probing and frozen-backbone cross-sensor gains over direct feature-alignment baselines disappear; if they do not, the architecture-gap account is not what is driving the results.

Watch

Extended reading notes

Core claim

Closing the cross-architecture gap—by coupling a LiDAR backbone with a student ViT from the same family as a frozen VFM teacher and supervising over patch-token representations built by Frustum Pooling and Frustum Attention with visibility masking—produces LiDAR features that transfer better under frozen-backbone and lightweight-head protocols than direct heterogeneous VFM-to-3D distillation.

Load-bearing premise

The method assumes that accurate, synchronized multi-camera–LiDAR calibrations give trustworthy frustum–patch correspondences, so aligning those tokens to the image teacher is the right supervisory signal for point-wise semantics.

Editorial extensions

If this is right

  • Frozen LiDAR backbones plus a small head become more competitive for new platforms when pretraining used same-family ViT token supervision rather than direct cross-architecture regression.
  • Adding more unlabeled multi-sensor frames continues to improve linear-probing averages under TOLiD-style coupling more than under prior feature-level alignment that saturates earlier.
  • Cross-sensor unsupervised domain adaptation (source-labeled head, frozen backbone, no target self-training) can reach or exceed methods that rely on target-domain self-training.
  • Deployment-time robustness to weather and sensor-failure corruptions improves in the frozen-representation regime without requiring full finetuning.
  • LiDAR-only inference remains possible because masked lifting reconstructs per-point features from multi-view tokens without inventing semantics from empty patches.

Reading between the lines

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

  • If architecture compatibility is the main lever, similar student–teacher family matching may help other sparse-to-dense transfers (radar, event cameras) without inventing new 3D foundation models from scratch.
  • Fleets that accumulate unlabeled multi-camera–LiDAR logs could treat continual TOLiD-style pretraining as a substitute for repeated dense 3D annotation campaigns when only a light head is trained per robot.
  • The method’s dependence on calibrated frustums suggests a natural stress test: whether learned or approximate cross-modal geometry can replace factory extrinsics without losing the scaling advantage.
  • When full finetuning equalizes methods under clean data, the practical value of the approach is concentrated in few-label, frozen-backbone, and corruption settings rather than fully supervised in-domain training.
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. TOLiD proposes a self-supervised VFM-to-LiDAR pretraining pipeline that couples a LiDAR backbone with a student ViT (same family as a frozen DINOv2 teacher) so that distillation occurs over compatible patch tokens rather than forcing a heterogeneous 3D encoder to match frozen image embeddings. Point features inside each image-patch frustum are mapped to tokens via Frustum Pooling (stage 0) and Frustum Attention (deeper stages); visibility-masked cosine+smooth-ℓ1 losses align student [CLS] and patch tokens to the teacher; for LiDAR-only use, tokens are lifted back to points with masked bilinear sampling and concatenated with backbone features. The method is pretrained on up to five heterogeneous camera–LiDAR datasets and evaluated under linear probing, label-efficient and full finetuning, four frozen-backbone UDA pairs, Robo3D corruptions, and data-scaling curves, reporting gains over multi-dataset baselines such as ScaLR, especially with frozen backbones and lightweight heads.

Significance. If the architecture-gap diagnosis is correct and the gains are not primarily capacity or residual-VFM artifacts, the work is a meaningful advance for label-efficient 3D perception: it shows that multi-sensor pretraining can be made more scalable and that frozen LiDAR representations can transfer across sensors and corruptions without target self-training. Strengths include broad multi-dataset evaluation (nuScenes, SemanticKITTI, PandaSet-64/GT, Waymo), explicit scaling curves (169k→342k frames), four UDA pairs under a frozen-backbone protocol, Robo3D robustness, and concrete modules (Frustum Attention, masked bilinear sampling) that are clearly specified. The practical emphasis on deployment with lightweight heads is well motivated for robotics fleets that accumulate unlabeled multi-sensor data.

major comments (3)
  1. [§I, §III.B–F, Tab. V, Fig. 1] Central claim isolation (§I, §III intro, Abstract): Gains are attributed to supervising compatible patch-token representations rather than heterogeneous VFM o3D distillation. At evaluation, however, LP/UDA/finetuning use hybrid features [f_i ∥ r_i] where r_i is the masked-bilinear lift of multi-stage student-ViT tokens (Eq. 10, §III.F; Fig. 1). The student is a full DINOv2 ViT-B/14 with J-stage Frustum injections, so capacity and residual teacher-family pathway are confounded with the architecture-gap story. Tab. V only swaps pooling vs attention and masked vs naive sampling under a small DINO-S setting; it does not remove the student ViT, freeze it after init, or match parameter/FLOP budgets against ScaLR’s standalone 3D encoder. A load-bearing ablation (e.g., L_θ-only features after distillation; capacity-matched pure-3D student; or student frozen after init) is needed to support the t
  2. [Tab. I, Tab. II, §IV.B, §III.F] What is frozen and compared (Tab. I–IV, §IV.B): Linear probing and UDA are described as “frozen backbone + lightweight head,” but it is unclear whether G_φ remains active (and how many parameters) versus L_θ alone. ScaLR and prior baselines are heterogeneous 3D encoders without a residual ViT pathway. Please report parameter counts, inference FLOPs, and an L_θ-only LP/UDA column so that +3.9 mIoU (nuScenes LP) and the UDA averages can be interpreted fairly. Without this, the SOTA frozen-backbone claim is not fully comparable.
  3. [Tab. I, Tab. IV, Abstract, §VI] Full-finetuning and saturation (Tab. I, Tab. IV): Under 100% nuScenes labels TOLiD (77.3) is slightly below ScaLR (78.4); under full finetuning on Robo3D both methods converge (~65.7). The paper’s strongest evidence is therefore frozen/LP transfer. This should be stated more carefully in the abstract and conclusion so the architecture-gap claim is not over-generalized to all regimes. Discuss whether the hybrid pathway mainly improves linear separability of frozen features rather than asymptotic supervised accuracy.
minor comments (5)
  1. [§III.B–C] Notation: W_p / W_p^{(j)} and stage index j are introduced with occasional omissions (§III.C); keep stage superscripts consistent in Eqs. (2)–(5).
  2. [Abstract, §I, §V] Typographic: “TOLiD” is often concatenated to the next word (e.g., “proposeTOLiD”, “evaluateTOLiDon”); insert spaces throughout.
  3. [§IV.C] Fixed-camera protocol (§IV.C): replacing per-frame calibrations with a dataset-level rig (including virtual rear cameras on Waymo) is pragmatic but may inject systematic projection error; a short sensitivity note would help.
  4. [Fig. 4, Tab. III] Fig. 4 caption and Tab. III: average LP gap +2.0 o+4.2 is useful; also report per-dataset variance or confidence intervals if multiple seeds were run.
  5. [§II.B] Related work: briefly contrast with concurrent hybrid 2D–3D adapters (e.g., ViT adapters for LiDAR) so the novelty of frustum-token injection vs. feature regression is sharper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: empirical VFM-to-LiDAR method evaluated on external public benchmarks; losses and modules are not algebraic restatements of reported mIoU.

full rationale

TOLiD is a self-supervised pretraining architecture paper. Its load-bearing claims are empirical transfer metrics (linear probing, few-shot finetuning, frozen-backbone UDA, Robo3D robustness) on public datasets (nuScenes, SemanticKITTI, PandaSet, Waymo, Robo3D) against independently published baselines (ScaLR, Seal, MuDDoS, etc.). The training objectives (visibility-masked L_cls + L_patch over DINOv2 teacher tokens; Eqs. 6–9) and the geometric modules (Frustum Pooling Eq. 2, Frustum Attention Eqs. 3–5, Masked Bilinear Sampling Eq. 10) are design choices whose success is measured by a different quantity (point-wise mIoU under frozen or finetuned heads). Nothing in the chain defines the reported mIoU in terms of the distillation loss, fits a parameter on a subset and renames a related quantity as a prediction, or imports a uniqueness theorem from overlapping authors. Self-citations (e.g., Point-PNG, Pair-VPR) are peripheral and not load-bearing for Tables I–V. Capacity confounds raised by the skeptic (student ViT residual pathway vs pure 3D encoder) are experimental-isolation concerns, not circularity under the stated patterns. Score 0 is therefore the correct honest finding.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claim rests on standard multi-view geometry, the transferability of DINOv2 patch tokens as distillation targets, and several engineering choices (backbone, fusion stages, loss mix) rather than on free physical constants. Invented modules are algorithmic constructs validated only inside this paper’s experiments. No machine-checked theory; evidence is empirical mIoU tables.

free parameters (4)
  • AdamW initial LR and schedule = 5e-4 → 1e-6, 25 epochs
    5e-4 cosine-decayed to 1e-6 with 10% warmup and weight decay 0.03; chosen by authors, not derived.
  • Frustum Attention insertion interval and stage count J = every 4th layer
    FA every 4th ViT layer; multi-stage integration count J is a design hyperparameter controlling how often LiDAR features are injected.
  • Distillation loss mix and teacher-token EMA standardization = 0.5 Lcls + 0.5 Lpatch; β 1.0→0.001
    Equal weight on cosine + smooth-L1 for CLS and masked patches; EMA momentum cosine from 1.0 to 0.001 following cited practice—hand-set training recipe.
  • Student/teacher ViT and LiDAR backbone sizes = WI-48 / ViT-B / ViT-L
    WaffleIron-48 (768-d), student ViT-B/14, teacher ViT-L/14 are capacity choices that affect reported mIoU.
assumptions (4)
  • domain assumption Perspective projection with known intrinsics/extrinsics yields correct point–patch frustum membership for distillation and lifting.
    Section III.A, Eq. 1; entire pipeline depends on calibrated multi-sensor geometry.
  • domain assumption DINOv2 teacher patch and CLS tokens are useful semantic targets for outdoor LiDAR point features under visibility masking.
    Sections III.E and IV.D; teacher is frozen DINOv2 without task-specific proof that its space is optimal for 3D segmentation.
  • ad hoc to paper Mean pooling (stage 0) plus residual cross-attention (deeper stages) is a sufficient point-set → token map for variable-density LiDAR.
    Sections III.B–III.C, Eqs. 2–5; design choice of TOLiD, only weakly ablated.
  • standard math Linear algebra and softmax attention behave as in standard deep learning (no new math claimed).
    Used throughout Frustum Attention and losses.
invented entities (2)
  • Frustum Attention (content-aware 3D→2D cross-attention per patch frustum)
    purpose: Selectively aggregate variable LiDAR points into ViT tokens conditioned on current token state.
    Introduced in §III.C as a TOLiD-specific module; only validated by small ablations in this paper.
  • Masked Bilinear Sampling for 2D→3D token lifting
    purpose: Lift ViT tokens to points while suppressing empty-patch contamination.
    §III.F / Eq. 10; engineering construct without external independent validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TOLiD: Bridging the Architecture Gap in Vision Foundation Model to LiDAR Pretraining via Token Lifting for Distillation." pith.science (2026). https://pith.science/paper/UKMUKT2I

@misc{pith2026260710762,
  author       = {Pith},
  title        = {Pith review of: TOLiD: Bridging the Architecture Gap in Vision Foundation Model to LiDAR Pretraining via Token Lifting for Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UKMUKT2I}},
  note         = {Machine review of arXiv:2607.10762}
}
read the original abstract

Cross-modal distillation from Vision Foundation Models (VFMs) to LiDAR backbones has recently emerged as a self-supervised pretraining strategy that reduces reliance on dense point-wise annotation for 3D scene understanding. However, existing distillation pipelines typically treat the VFM as a frozen feature source and train a heterogeneous 3D backbone to match fixed image embeddings, forcing the student to bridge both the modality gap and the cross-architecture gap between dense ViT token representations and sparse 3D encoders. We propose TOLiD, a self-supervised pretraining method for LiDAR representation learning that addresses this gap by coupling a LiDAR backbone with a student Vision Transformer (ViT) initialized from a frozen VFM teacher and applying supervision over compatible patch-token representations. TOLiD converts the set of point features within each image patch frustum into a token using Frustum Pooling followed by Frustum Attention, and performs token-level distillation with visibility masking. For LiDAR-only deployment, we lift token features back to per-point representations using masked bilinear sampling to avoid patches that have limited LiDAR points. We extensively evaluate TOLiD on five heterogeneous LiDAR datasets and four cross-sensor adaptation pairs, demonstrating improved transfer with frozen backbones and lightweight heads.

Figures

Figures reproduced from arXiv: 2607.10762 by the authors.

Figure 1
Figure 1. Overview of the TOLiD. During pretraining (top), Lθ converts the point features into patch tokens via Frustum Pooling (FP) and Frustum Attention (FA) at J stages and integrates into the Gϕ; visibility-masked distillation aligns them with the frozen teacher T . During finetuning (bottom), token features are lifted back to per-point representations via Masked Bilinear Sampling (MBS) across all views and concatenated w… view at source ↗
Figure 3
Figure 3. Masked Bilinear Sampling for 2D→3D Lifting. Empty patches (red) are suppressed by the occupancy mask Mv n and the contributions from valid tokens (blue) are renor￾malised, preventing spurious features in lifted 3D points. Global Alignment: We align the [CLS] tokens through Lcls = Ldist  ψ(z S,v cls ), z T ,v cls  . (7) Visibility-Masked Local Alignment: Patch-level supervi￾sion is applied only to tokens supported … view at source ↗
Figure 4
Figure 4. Effect of scaling pretraining data. Average [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Qualitative results on SemanticKITTI (linear probing). TOLiD produces more semantically consistent predictions than ScaLR [11]. yields 3D embeddings that are more semantically structured and linearly separable than those obtained through cross￾architectural distillatio…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 1 linked inside Pith

  1. [1]

    Semantickitti: A dataset for seman- tic scene understanding of lidar sequences,

    J. Behley, M. Garbade,et al., “Semantickitti: A dataset for seman- tic scene understanding of lidar sequences,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 9297–9307

    2019

  2. [2]

    WildScenes: A Benchmark for 2D and 3D Semantic Segmentation in Large-scale Natural Environ- ments,

    K. Vidanapathirana, J. Knights,et al., “WildScenes: A Benchmark for 2D and 3D Semantic Segmentation in Large-scale Natural Environ- ments,”The International Journal of Robotics Research, vol. 44, no. 4, pp. 532–549, 2025

    2025

  3. [3]

    Deep evidential uncertainty esti- mation for semantic segmentation under out-of-distribution obstacles,

    S. Ancha, P. R. Osteen, and N. Roy, “Deep evidential uncertainty esti- mation for semantic segmentation under out-of-distribution obstacles,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 6943–6951

    2024

  4. [4]

    Suma++: Efficient lidar-based semantic slam,

    X. Chen, A. Milioto,et al., “Suma++: Efficient lidar-based semantic slam,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2019, pp. 4530–4537

    2019

  5. [5]

    Pointmoseg: Sparse tensor-based end-to- end moving-obstacle segmentation in 3-d lidar point clouds for au- tonomous driving,

    Y . Sun, W. Zuo,et al., “Pointmoseg: Sparse tensor-based end-to- end moving-obstacle segmentation in 3-d lidar point clouds for au- tonomous driving,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 510–517, 2020

    2020

  6. [6]

    Theia: Distilling diverse vision foundation models for robot learning,

    J. Shang, K. Schmeckpeper,et al., “Theia: Distilling diverse vision foundation models for robot learning,” in8th Annual Conference on Robot Learning, 2024

    2024

  7. [7]

    Pair-VPR: Place-Aware Pre-Training and Contrastive Pair Classification for Visual Place Recognition with Vision Transformers,

    S. Hausler and P. Moghadam, “Pair-VPR: Place-Aware Pre-Training and Contrastive Pair Classification for Visual Place Recognition with Vision Transformers,”IEEE Robotics and Automation Letters, 2025

    2025

  8. [8]

    ImLPR: Image-based LiDAR Place Recognition using Vision Foundation Models,

    M. Jung, L. F. T. Fu,et al., “ImLPR: Image-based LiDAR Place Recognition using Vision Foundation Models,” inProceedings of The 9th Conference on Robot Learning, vol. 305, 2025, pp. 3318–3340

    2025

Show all 36 references
  1. [9]

    Label-efficient lidar semantic segmen- tation with 2d-3d vision transformer adapters,

    J. Hindel, R. Mohan,et al., “Label-efficient lidar semantic segmen- tation with 2d-3d vision transformer adapters,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 4115–4122

    2025

  2. [10]

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

    Y .-C. Liu, Y .-K. Huang,et al., “Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining,”arXiv preprint arXiv:2104.04687, 2021

    2021 arXiv

  3. [11]

    Three pillars improving vision foundation model distillation for lidar,

    G. Puy, S. Gidaris,et al., “Three pillars improving vision foundation model distillation for lidar,” inCVPR, June 2024, pp. 21 519–21 529

    2024

  4. [12]

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

    C. Sautier, G. Puy,et al., “Image-to-lidar self-supervised distillation for autonomous driving data,” inCVPR, 2022, pp. 9891–9901

    2022

  5. [13]

    Segment any point cloud sequences by distilling vision foundation models,

    Y . Liu, L. Kong,et al., “Segment any point cloud sequences by distilling vision foundation models,”Advances in Neural Information Processing Systems, vol. 36, pp. 37 193–37 229, 2023

    2023

  6. [14]

    Cleverdistiller: Simple and spatially consistent cross-modal distillation,

    H. Govindarajan, M. K. Wozniak,et al., “Cleverdistiller: Simple and spatially consistent cross-modal distillation,” inBMVC, 2025

    2025

  7. [15]

    Cross-modal self-supervised learning with effective contrastive units for lidar point clouds,

    M. Cai, C. Luo,et al., “Cross-modal self-supervised learning with effective contrastive units for lidar point clouds,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 9468–9475

    2024

  8. [16]

    Point-PNG: Conditional Pseudo- Negatives Generation for Point Cloud Pre-Training,

    S. Mahendren, S. Rahman,et al., “Point-PNG: Conditional Pseudo- Negatives Generation for Point Cloud Pre-Training,”IEEE Access, vol. 13, pp. 208 612–208 625, 2025

    2025

  9. [17]

    Expert-enhanced masked point modeling for point cloud self-supervised learning,

    Y . Liu, Y . Zha,et al., “Expert-enhanced masked point modeling for point cloud self-supervised learning,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 1436– 1443

    2025

  10. [18]

    Clip2scene: Towards label-efficient 3d scene understanding by clip,

    R. Chen, Y . Liu,et al., “Clip2scene: Towards label-efficient 3d scene understanding by clip,” inCVPR, 2023, pp. 7020–7030

    2023

  11. [19]

    Largead: Large-scale cross-sensor data pretraining for autonomous driving,

    L. Kong, X. Xu,et al., “Largead: Large-scale cross-sensor data pretraining for autonomous driving,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  12. [20]

    Do vision transformers see like convolutional neural networks?

    M. Raghu, T. Unterthiner,et al., “Do vision transformers see like convolutional neural networks?” inAdvances in Neural Information Processing Systems, vol. 34, 2021, pp. 12 116–12 128

    2021

  13. [21]

    One-for-all: Bridge the gap between hetero- geneous architectures in knowledge distillation,

    Z. Hao, J. Guo,et al., “One-for-all: Bridge the gap between hetero- geneous architectures in knowledge distillation,”Advances in Neural Information Processing Systems, vol. 36, pp. 79 570–79 582, 2023

    2023

  14. [22]

    Cumulative spatial knowledge distillation for vision transformers,

    B. Zhao, R. Song, and J. Liang, “Cumulative spatial knowledge distillation for vision transformers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 6146–6155

    2023

  15. [23]

    nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti,et al., “nuscenes: A multimodal dataset for autonomous driving,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 621–11 631

    2020

  16. [24]

    Pandaset: Advanced sensor suite dataset for autonomous driving,

    P. Xiao, Z. Shao,et al., “Pandaset: Advanced sensor suite dataset for autonomous driving,” in2021 IEEE International Intelligent Trans- portation Systems Conference (ITSC), 2021, p. 3095–3101

    2021

  17. [25]

    Scalability in perception for au- tonomous driving: Waymo open dataset,

    P. Sun, H. Kretzschmar,et al., “Scalability in perception for au- tonomous driving: Waymo open dataset,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2446–2454

    2020

  18. [26]

    Improving multimodal distillation for 3d semantic segmentation under domain shift,

    B. Michele, A. Boulch,et al., “Improving multimodal distillation for 3d semantic segmentation under domain shift,” in36th British Machine Vision Conference 2025, BMVC 2025, Sheffield, UK, November 24-27,

    2025

  19. [27]

    Self-supervised image-to-point distil- lation via semantically tolerant contrastive loss,

    A. Mahmoud, J. S. Hu,et al., “Self-supervised image-to-point distil- lation via semantically tolerant contrastive loss,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7102–7110

    2023

  20. [28]

    Minimal-entropy correlation alignment for unsupervised deep domain adaptation,

    P. Morerio, J. Cavazza, and V . Murino, “Minimal-entropy correlation alignment for unsupervised deep domain adaptation,” inICLR, 2018

    2018

  21. [29]

    Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation,

    C. Saltori, F. Galasso,et al., “Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation,” inEuropean conference on computer vision. Springer, 2022, pp. 586–602

    2022

  22. [30]

    Exploiting the complementarity of 2d and 3d networks to address domain-shift in 3d semantic seg- mentation,

    A. Cardace, P. Z. Ramirez,et al., “Exploiting the complementarity of 2d and 3d networks to address domain-shift in 3d semantic seg- mentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2023, pp. 98–109

    2023

  23. [31]

    Learning to adapt sam for segmenting cross- domain point clouds,

    X. Peng, R. Chen,et al., “Learning to adapt sam for segmenting cross- domain point clouds,” inEuropean Conference on Computer Vision, 2024, pp. 54–71

    2024

  24. [32]

    Using a waffle iron for automotive point cloud semantic segmentation,

    G. Puy, A. Boulch, and R. Marlet, “Using a waffle iron for automotive point cloud semantic segmentation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp. 3379–3389

    2023

  25. [33]

    DINOv2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet,et al., “DINOv2: Learning robust visual features without supervision,”Transactions on Machine Learning Research, 2024

    2024

  26. [34]

    Dune: Distilling a universal encoder from heterogeneous 2d and 3d teachers,

    M. B. Sarıyıldız, P. Weinzaepfel,et al., “Dune: Distilling a universal encoder from heterogeneous 2d and 3d teachers,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 30 084–30 094

    2025

  27. [35]

    Robo3D: Towards robust and reliable 3d perception against corruptions,

    L. Kong, Y . Liu,et al., “Robo3D: Towards robust and reliable 3d perception against corruptions,” inICCV, 2023, pp. 19 994–20 006

    2023

  28. [36]

    4d contrastive superflows are dense 3d representation learners,

    X. Xu, L. Kong,et al., “4d contrastive superflows are dense 3d representation learners,” inEuropean Conference on Computer Vision, 2024, pp. 58–80

    2024

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.