Pith. sign in

REVIEW 4 major objections 4 minor 53 references

Vernata: Self-Supervised Learning of LiDAR Point Representations

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Vernata, a three-part extension of the Sonata self-supervised framework, lifts LiDAR semantic segmentation mIoU by 5.9–7.3 points on outdoor data.

desk verdict Solid incremental extension of Sonata with clean ablations, but single-run numbers, a misleading reduced-modality abstract, and an unexamined occlusion problem in cross-modal distillation keep the headline gains soft. read the letter →

arxiv 2608.06919 v1 pith:4C3VF6SL submitted 2026-08-07 cs.CV cs.RO

Oliver Lemke , Alexander Liniger , Abel Gawel , Marco Hutter This is my paper · ORCID
classification cs.CVcs.RO
keywords self-supervisedlearningLiDARpointcloudscross-modaldistillationsemanticsegmentationmemorybanksparseviewaugmentationSonataDINOv2
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 claims that a self-supervised learner for outdoor LiDAR point clouds can be made substantially more effective by augmenting the Sonata architecture with three components: sparse view augmentation, a Sinkhorn-Knopp memory bank, and cross-modal distillation from dense 2D image features. On the TartanGround and Waymo datasets, these extensions raise linear-probe semantic segmentation mIoU from 48.8 to 54.7 and from 49.8 to 57.1, respectively, with cross-modal distillation delivering the largest single gain. The motivation is practical: 3D annotation is expensive, so general-purpose representations that transfer to downstream tasks with minimal labels are valuable for robotics. If the claim holds, the method offers a recipe for learning useful LiDAR features without any human labels, even when color and surface normals are absent.

What carries the argument

The framework's load-bearing mechanism is multi-teacher distillation: a 3D teacher (EMA of the student) provides dense prototype targets via Sinkhorn-Knopp normalization, and a frozen 2D teacher (DINOv2 features upsampled with LoftUp and backprojected via known camera intrinsics and extrinsics) provides per-point cosine-similarity targets. The student, fed masked, local, and sparsely subsampled views, must match both teachers through a sum of a uni-modal cross-entropy loss and a cross-modal cosine loss. The sparse-view objective enforces density invariance, the memory bank decouples normalization statistics from batch size, and the cross-modal objective injects fine-grained semantic signal that pure 3D SSL lacks, especially at long range.

What would settle it

A decisive test would be to rerun the full pretraining pipeline with the cross-modal distillation target replaced by random or inverted features (for example, shuffled point-to-pixel assignments). If the mIoU gains over Sonata persist, the semantic correspondence that the paper claims is load-bearing is not actually responsible for the improvement, and the central explanation collapses.

Watch

Extended reading notes

Core claim

Vernata is a multi-modal, multi-teacher distillation framework that extends the Sonata self-supervised point-cloud learner to outdoor LiDAR. The student network sees masked, local, and sparse-subsampled views of a scene and must match the prototype assignments of an EMA-updated 3D teacher, while also aligning its point features to dense, high-resolution DINOv2 image features that have been upsampled with LoftUp and backprojected onto the points. A FIFO memory bank of prototype scores stabilizes Sinkhorn-Knopp normalization when training on only four GPUs. The authors report that this combination consistently outperforms both the frozen ScanNet-pretrained Sonata checkpoint and a Sonata variant finetuned on the target LiDAR data, with mIoU gains of +5.9 points on TartanGround and +7.3 points on Waymo, and that the pretrained representations remain competitive when color and normals are removed at inference.

Load-bearing premise

The backprojected DINOv2 features remain semantically meaningful for each 3D point after LoftUp upsampling and spatial assignment, so the cross-modal distillation loss injects true semantic guidance rather than noise.

Editorial extensions

If this is right

  • If the reported gains hold, self-supervised pretraining on unlabeled outdoor LiDAR can yield semantic segmentation features that rival or beat fully supervised training in low-label regimes, shown by a +2.5 mIoU edge over PTv3 on a 220-frame custom dataset.
  • The sparse-view augmentation suggests that density-invariant representations are a key ingredient for outdoor LiDAR SSL, so other point-cloud tasks such as object detection or panoptic segmentation are plausible beneficiaries.
  • The memory bank mechanism offers a practical path to train SSL models on limited GPU budgets without sacrificing representation quality, widening access to this type of pretraining.
  • Cross-modal distillation from frozen 2D foundation models appears to be the single most valuable extension, implying that high-resolution image features carry semantic information that point-only SSL does not capture.
  • Reduced-modality performance (coordinates only) indicates that Vernata's representations are robust when color and normals are unavailable, a common constraint on real robots.

Reading between the lines

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

  • If cross-modal distillation's value comes from dense high-resolution features rather than the specific DINOv2 teacher, then distilling from other strong 2D encoders could yield further gains; the paper's Table V result that dense matching helps at long range makes this a direct, testable next step.
  • The paper does not quantify sensitivity to camera-LiDAR calibration error; a plausible extension is to perturb extrinsics during training or evaluation to measure robustness of the distilled representations, which would inform deployment on robots with imperfect calibration.
  • Because the memory bank decouples normalization from batch size, one could plausibly train with even smaller batches or on longer sequences by increasing bank size; the relationship between bank size and representation quality at different GPU counts is left uncharacterized.
  • The TartanGround protocol introduced here (class mapping, split, and scales) could become a standard evaluation for outdoor LiDAR SSL, but its stability across random seeds and hyperparameter changes is not yet established.
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 / 4 minor

Summary. The paper proposes Vernata, a self-supervised learning framework for outdoor LiDAR point clouds that extends the Sonata architecture with three components: sparse view augmentation to combat range-dependent density variation, a Sinkhorn-Knopp memory bank to stabilize small-batch training, and cross-modal distillation from frozen DINOv2 image features. The method is evaluated by linear probing for semantic segmentation on TartanGround, Waymo, and a small in-house dataset, reporting mIoU gains of +5.9 on TartanGround and +7.3 on Waymo over a Sonata baseline that is fine-tuned on the same target data. Ablations show the largest contribution comes from cross-modal distillation, and additional experiments examine robustness to missing color/normals and comparison with a patch-based distillation baseline.

Significance. If the results hold, the paper makes a useful empirical contribution: it transfers a strong indoor SSL point-cloud framework to outdoor LiDAR, identifies density variation as a key challenge, and demonstrates a concrete way to inject 2D vision foundation model semantics. Strengths include the controlled Sonata-finetuned baseline, component-wise ablations, evaluation on Waymo as an independent established benchmark, a released public implementation, and a released evaluation protocol for TartanGround. The reported gains are sizable, especially on Waymo. However, the paper's central claims rely on single-run experiments and on a cross-modal distillation objective whose target assignment is not validated for occlusions; these issues need to be addressed before the claims can be considered fully supported.

major comments (4)
  1. [Section III-D, Eq. (3)] The cross-modal loss samples a 2D feature for every LiDAR point from its projected image coordinate, but no visibility or occlusion test is specified. For any LiDAR point that is behind an occluding surface in a given camera view—common in outdoor scenes, e.g., ground points behind a vehicle—the bilinearly sampled DINOv2 feature belongs to the occluder rather than to the measured point, so L_sim systematically pushes the student toward incorrect targets. Since Table II attributes the largest single gain to CMD (+4.6 on TartanGround, +6.7 on Waymo), this is a load-bearing assumption. Please add a depth-based visibility mask or otherwise exclude occluded assignments, report the resulting mIoU, and include a sensitivity analysis to camera-LiDAR extrinsic calibration error.
  2. [Tables I and II] All experimental results are single-run point estimates without standard deviations or significance tests. In Table II, the difference between the SP row and the SP+MB row is only +0.8/+0.9 mIoU, which is likely within run-to-run variability for semantic segmentation benchmarks; even the headline +5.9/+7.3 gains need variance estimates before the word "significant" is used. Please report mean and standard deviation over at least three random seeds for the main comparisons and ablations, or provide a fixed-seed protocol with the released code that makes the results exactly reproducible.
  3. [Section IV, TartanGround protocol] The TartanGround evaluation is a self-defined protocol: 3-frame accumulation, mapping of 1496 raw labels to 7 classes, a stratified split, and a curated 6501-sample linear-probe subset. The headline +5.9 mIoU result therefore depends on choices that are not independently established. Please provide the full class mapping and class distributions, the exact split and curation criterion, and a robustness check such as an alternative split or k-fold evaluation; the released benchmark should include these details so the result can be reproduced and compared fairly.
  4. [Section IV-D and abstract] The reduced-modality experiments in Table IV are conducted with the "CMD-less variant" because cross-modal distillation was omitted "due to the additional training overhead," yet the abstract states generally that "the self-supervised approach maintains strong performance even in reduced-modality settings (lacking color or normals)." The full Vernata method is never tested without color/normals, so the robustness claim is only established for the variant. Please either evaluate the full model under reduced input modalities (CMD is a pretraining-only objective and should be compatible with coordinate-only inference) or reword the claims so they explicitly refer to the CMD-less variant.
minor comments (4)
  1. [Fig. 5 caption vs. Table III] The Figure 5 caption says the linear head is fine-tuned on an in-house dataset of 311 frames, while Table III reports n=220 for the custom dataset; please clarify the discrepancy.
  2. [Section III-B] The sparse view augmentation subsamples the point cloud uniformly at random, but real LiDAR density varies by range and beam pattern; a sentence explaining why uniform subsampling is a sufficient surrogate for the actual sparsity distribution would improve clarity.
  3. [Table II] The small mAcc dip on Waymo in the full model is attributed to a "class-balancing trade-off," but no class-wise results are provided to support this explanation; please either add per-class numbers or soften the attribution.
  4. [Eq. (2)] It would help to state explicitly that the memory bank stores per-point prototype scores from previous batches and to specify whether the bank is reset between training runs or epochs.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported gains are measured on held-out labels against external Sonata baselines, and no equation or fitted parameter reduces a prediction to an input.

full rationale

The derivation chain is self-contained and externally grounded. The three contributions (sparse views, memory bank, cross-modal distillation) are training objectives whose effect is measured by linear probing on held-out semantic labels; the reported mIoU numbers are empirical outcomes, not quantities defined by the method's equations. The cross-modal loss in Eq. (3) aligns student features with features from a frozen external DINOv2 teacher, so the target h_j^(t) is not constructed from the student's output or from a fitted parameter of this paper. No equation equates a prediction with a training target by construction. The TartanGround evaluation protocol is defined by the authors, but this is a benchmark-selection choice rather than a circular derivation, and the Waymo results provide an independent public benchmark with consistent gains. Citations to Sonata, ScaLR, LoftUp, and DINOv2 are to external prior work, not to the present authors' own unverified claims; the only author-overlapping citations are dataset references (GrandTour, TartanGround), which are not load-bearing for the method's derivation. The occlusion-sensitivity concern about backprojection is a correctness or robustness risk, not circularity, because it does not make any reported result true by definition.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard SSL machinery from Sonata, SwAV, and DINOv2, plus several hand-chosen benchmark and training choices. No new physical or conceptual entities are postulated. The largest uncertainties are the custom TartanGround protocol, single-run results without error bars, and reliance on accurate camera-LiDAR alignment for the cross-modal objective.

free parameters (6)
  • Sparse view subsampling ratio schedule = (0.9,1.0) decreasing to (0.5,0.7)
    Hand-chosen progressive schedule for sparse-to-dense augmentation; controls strength of density randomization in Section III-B.
  • Memory bank size = 50,000 (GrandTour/TartanGround), 100,000 (Waymo)
    Chosen per dataset to stabilize Sinkhorn-Knopp normalization under 4-GPU training; affects prototype statistics.
  • View and mask schedules = global (0.7,1.0)->(0.4,1.0); local (0.1,0.4); mask size 0.1->0.4, ratio 0.3->0.7
    Standard Sonata-style view generation hyperparameters used for all variants, from Table VI.
  • TartanGround preprocessing = 3-frame accumulation; 50m x 50m crop; 0.1m voxel grid
    Hand-selected evaluation preprocessing; affects point density and class distribution for the custom benchmark.
  • TartanGround curated linear-probe subset = 6,501 samples
    Authors restrict linear probing to a subset of 'highest-quality' labels; this curation can affect reported mIoU.
  • Training steps and learning rates = SSL: 20,000 steps, LR 2e-4; LP: 10,000/20,000 steps, LR 1e-3
    Hand-chosen optimization schedule from Table VI; not swept.
assumptions (5)
  • domain assumption ScanNet-pretrained Sonata checkpoint is a valid initialization for outdoor LiDAR SSL.
    All models are initialized from this checkpoint (Section IV, Implementation Details); if it carries indoor biases, the reported gains could partly reflect adaptation rather than the proposed extensions.
  • domain assumption Backprojected DINOv2 features provide semantically meaningful targets for 3D points.
    Cross-modal distillation (Section III-D) assumes frozen 2D VFM features remain meaningful after projection and upsampling; the ablation supports this but does not prove calibration correctness.
  • domain assumption Sinkhorn-Knopp with a memory bank preserves accurate prototype assignments at small batch sizes.
    Memory bank mechanism (Eq. 2) assumes stale prototype scores in a FIFO queue still yield a good global distribution estimate.
  • domain assumption Uniform random subsampling of a dense view approximates range-dependent LiDAR sparsity.
    Sparse view augmentation (Section III-B) assumes this simulation transfers to real density variation; if real sparsity has structured patterns, the augmentation may not generalize.
  • domain assumption The custom TartanGround protocol (1496 labels to 7 classes, split, curated subset) measures semantic segmentation quality fairly.
    Because TartanGround lacked a protocol, the authors define one; the resulting mIoU numbers are only as meaningful as this mapping and split.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vernata: Self-Supervised Learning of LiDAR Point Representations." pith.science (2026). https://pith.science/paper/4C3VF6SL

@misc{pith2026260806919,
  author       = {Pith},
  title        = {Pith review of: Vernata: Self-Supervised Learning of LiDAR Point Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4C3VF6SL}},
  note         = {Machine review of arXiv:2608.06919}
}
read the original abstract

LiDAR serves as a primary sensing modality for robots operating in outdoor environments. However, the performance of deep learning models in this domain is severely limited by the scarcity of labeled data, a direct result of the high cost of 3D annotation. Self-supervised learning addresses this scarcity by learning general-purpose features from unlabeled data. In this work, we present a multi-modal, multi-teacher distillation framework for self-supervised learning on outdoor LiDAR point clouds. Building upon the Sonata architecture, we introduce Vernata, consisting of three extensions: sparse view augmentation to improve robustness against varying point densities, a memory bank mechanism to stabilize resource-constrained training, and cross-modal distillation utilizing dense, high-resolution 2D image features to enable fine-grained semantic guidance. We evaluate our method on the GrandTour, TartanGround, and Waymo datasets, as well as data collected from our own robotic platforms. Our experiments demonstrate a significant performance improvement over Sonata baselines, yielding mIoU scores of 54.7 on TartanGround (+5.9 points, +12.1%) and 57.1 on Waymo (+7.3 points, +14.7%). Finally, we show that the self-supervised approach maintains strong performance even in reduced-modality settings (lacking color or normals), achieving competitive mIoU scores of 49.4 and 50.2 on the respective datasets.

Figures

Figures reproduced from arXiv: 2608.06919 by the authors.

Figure 1
Figure 1. Visualization of PCA Features in LiDAR Scenes. We visualize the principal components of point representations produced by Sonata [31] finetuned on LiDAR datasets, as well as our approach. Sonata struggles with separation between distinct semantic classes, as well as consistency across the inherent density variations of LiDAR point clouds. Our method integrates cross-modal distillation and a sparse-to-dense objective… view at source ↗
Figure 2
Figure 2. Overview of the Framework. We present a multi-modal multi-teacher distillation architecture for self-supervised point cloud learning. Building upon the Sonata [31] framework, highlighted in gray, our model is anchored by two teachers processing global views of the scene: A 3D teacher (top), employing a PTv3 [33] encoder and updated via EMA, and a frozen 2D teacher (bottom), producing DINOv2-S [20] features, upsample… view at source ↗
Figure 3
Figure 3. Sparsity Encoding Phenomenon. We show cosine similarity maps of a scene (left) for a reference point close (middle) and far away (right) from the robot. Despite the uniform terrain at both references, the embeddings diverge. Although our method (bottom) demonstrates improved consistency, particularly for distant points, the distinction between dense and sparse regions remains clearly visible. Reference highlighted w… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative Results. We compare qualitative segmentation results against the finetuned Sonata variant on TartanGround (top) and Waymo (bottom). In TartanGround, Vernata demonstrates significant improvements in segmenting challenging classes (stairs), alongside better p…
Figure 5
Figure 5. Figure 5: Inference on Real-World Trajectories. We present inference results for our CMD-less model variant, pretrained on GrandTour and TartanGround, with a linear head finetuned on an in-house dataset of just 311 frames. The top row displays inference on the GrandTour Triemli …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 29 canonical work pages

  1. [1]

    Autonomous spot: Long-range autonomous exploration of extreme environments with legged locomotion,

    A. Bouman, M. F. Ginting, N. Alatur, M. Palieri, D. D. Fan, T. Touma, T. Pailevanian, S.-K. Kim, K. Otsu, J. Burdick, and A.-a. Agha- Mohammadi, “Autonomous spot: Long-range autonomous exploration of extreme environments with legged locomotion,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, pp. 2518–2525

    work page 2020

  2. [2]

    Building a credible case for safety: Waymo’s approach for the determination of absence of unreasonable risk,

    F. Favaro, L. Fraade-Blanar, S. Schnelle, T. Victor, M. Pe ˜na, J. Engstrom, J. Scanlon, K. Kusano, and D. Smith, “Building a credible case for safety: Waymo’s approach for the determination of absence of unreasonable risk,” 2023. [Online]. Available: https://arxiv.org/abs/2306.01917

    arXiv 2023

  3. [3]

    Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning,

    Q. Gu, A. Kuwajerwala, S. Morin, K. M. Jatavallabhula, B. Sen, A. Agarwal, C. Rivera, W. Paul, K. Ellis, R. Chellappaet al., “Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 5021–5028

    2024

  4. [4]

    Visual language maps for robot navigation,

    C. Huang, O. Mees, A. Zeng, and W. Burgard, “Visual language maps for robot navigation,”arXiv preprint arXiv:2210.05714, 2022

    arXiv 2022

  5. [5]

    Aligning cyber space with physical world: A comprehensive survey on embodied ai,

    Y . Liu, W. Chen, Y . Bai, X. Liang, G. Li, W. Gao, and L. Lin, “Aligning cyber space with physical world: A comprehensive survey on embodied ai,”IEEE/ASME Transactions on Mechatronics, 2025

    2025

  6. [6]

    A survey on vision-language-action models for embodied ai,

    Y . Ma, Z. Song, Y . Zhuang, J. Hao, and I. King, “A survey on vision-language-action models for embodied ai,”arXiv preprint arXiv:2405.14093, 2024

    arXiv 2024

  7. [7]

    Spot-compose: A framework for open-vocabulary object retrieval and drawer manipulation in point clouds,

    O. Lemke, Z. Bauer, R. Zurbr ¨ugg, M. Pollefeys, F. Engelmann, and H. Blum, “Spot-compose: A framework for open-vocabulary object retrieval and drawer manipulation in point clouds,” in2nd Workshop on Mobile Manipulation and Embodied Intelligence at ICRA 2024, 2024

    work page 2024

  8. [8]

    Lost & found: Tracking changes from egocentric observations in 3d dynamic scene graphs,

    T. Behrens, R. Zurbr ¨ugg, M. Pollefeys, Z. Bauer, and H. Blum, “Lost & found: Tracking changes from egocentric observations in 3d dynamic scene graphs,”IEEE Robotics and Automation Letters, 2025

    work page 2025

Show all 53 references
  1. [9]

    Shan and C

    J. Shan and C. K. Toth,Topographic laser ranging and scanning: principles and processing. CRC press, 2018

    2018

  2. [10]

    Loam: Lidar odometry and mapping in real-time

    J. Zhang, S. Singhet al., “Loam: Lidar odometry and mapping in real-time.” inRobotics: Science and systems, vol. 2, no. 9. Berkeley, CA, 2014, pp. 1–9

    2014

  3. [11]

    Vilens: Visual, inertial, lidar, and leg odometry for all-terrain legged robots,

    D. Wisth, M. Camurri, and M. Fallon, “Vilens: Visual, inertial, lidar, and leg odometry for all-terrain legged robots,”IEEE Transactions on Robotics, vol. 39, no. 1, pp. 309–326, 2023

    2023

  4. [12]

    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. Caineet al., “Scalability in perception for autonomous driving: Waymo open dataset,” inCVPR, 2020, pp. 2446–2454

    2020

  5. [13]

    Grandtour: A legged robotics dataset in the wild for multi-modal perception and state estimation,

    J. Frey, T. Tuna, F. Fu, K. Patterson, T. Xu, M. Fallon, C. Ca- dena, and M. Hutter, “Grandtour: A legged robotics dataset in the wild for multi-modal perception and state estimation,”arXiv preprint arXiv:2602.18164, 2026

    2026 arXiv

  6. [14]

    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,” inCVPR, 2020, pp. 4604– 4612

    2020

  7. [15]

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

    Z. Zhuang, R. Li, K. Jia, Q. Wang, Y . Li, and M. Tan, “Perception- aware multi-sensor fusion for 3d lidar semantic segmentation,” in ICCV (ICCV), October 2021, pp. 16 280–16 290

    2021

  8. [16]

    Scaling laws for neural language models,

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei, “Scaling laws for neural language models,”arXiv preprint arXiv:2001.08361, 2020

    2001 arXiv

  9. [17]

    Explaining neural scaling laws,

    Y . Bahri, E. Dyer, J. Kaplan, J. Lee, and U. Sharma, “Explaining neural scaling laws,”Proceedings of the National Academy of Sciences, vol. 121, no. 27, p. e2311878121, 2024

    2024

  10. [18]

    Training compute-optimal large language models (2022),

    J. Hoffmann, S. Borgeaud, A. Mensch, E. Buchatskaya, T. Cai, E. Rutherford, D. de Las Casas, L. A. Hendricks, J. Welbl, A. Clark et al., “Training compute-optimal large language models (2022),” arXiv preprint arXiv:2203.15556, 2022

    2022 arXiv

  11. [19]

    Scaling laws of motion forecasting and planning–a technical report,

    M. Baniodeh, K. Goel, S. Ettinger, C. Fuertes, A. Seff, T. Shen, C. Gulino, C. Yang, G. Jerfel, D. Choeet al., “Scaling laws of motion forecasting and planning–a technical report,”arXiv preprint arXiv:2506.08228, 2025

    2025 arXiv

  12. [20]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. Haziza, F. Massa, A. El-Noubyet al., “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  13. [21]

    Sim ´eoni, H

    O. Sim ´eoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoaet al., “Dinov3,” arXiv preprint arXiv:2508.10104, 2025

    2025 arXiv

  14. [22]

    Sam 3: Segment anything with concepts,

    N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. Suris, C. Ryali, K. V . Alwala, H. Khedr, A. Huanget al., “Sam 3: Segment anything with concepts,”arXiv preprint arXiv:2511.16719, 2025

    2025 arXiv

  15. [23]

    Semantickitti: A dataset for semantic scene under- standing 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 under- standing of lidar sequences,” inICCV, 2019, pp. 9297–9307

    2019

  16. [24]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor scenes,” inCVPR, 2017, pp. 5828–5839

    2017

  17. [25]

    A simple frame- work for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple frame- work for contrastive learning of visual representations,” inInterna- tional conference on machine learning. PmLR, 2020, pp. 1597–1607

    2020

  18. [26]

    Towards a general-purpose foundation model for computational pathology,

    R. J. Chen, T. Ding, M. Y . Lu, D. F. Williamson, G. Jaume, A. H. Song, B. Chen, A. Zhang, D. Shao, M. Shabanet al., “Towards a general-purpose foundation model for computational pathology,” Nature medicine, vol. 30, no. 3, pp. 850–862, 2024

    2024

  19. [27]

    A foundation model for clinical-grade computational pathology and rare cancers detection,

    E. V orontsov, A. Bozkurt, A. Casson, G. Shaikovski, M. Zelechowski, K. Severson, E. Zimmermann, J. Hall, N. Tenenholtz, N. Fusiet al., “A foundation model for clinical-grade computational pathology and rare cancers detection,”Nature medicine, vol. 30, no. 10, pp. 2924–2935, 2024

    2024

  20. [28]

    Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery,

    Y . Cong, S. Khanna, C. Meng, P. Liu, E. Rozi, Y . He, M. Burke, D. Lobell, and S. Ermon, “Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery,”Advances in Neural Information Processing Systems, vol. 35, pp. 197–211, 2022

    2022

  21. [29]

    Exploring data-efficient 3d scene understanding with contrastive scene contexts,

    J. Hou, B. Graham, M. Nießner, and S. Xie, “Exploring data-efficient 3d scene understanding with contrastive scene contexts,” inCVPR, 2021, pp. 15 587–15 597

    2021

  22. [30]

    Masked scene contrast: A scalable framework for unsupervised 3d representation learning,

    X. Wu, X. Wen, X. Liu, and H. Zhao, “Masked scene contrast: A scalable framework for unsupervised 3d representation learning,” in CVPR, 2023, pp. 9415–9424

    2023

  23. [31]

    Sonata: Self-supervised learning of reliable point representations,

    X. Wu, D. DeTone, D. Frost, T. Shen, C. Xie, N. Yang, J. Engel, R. Newcombe, H. Zhao, and J. Straub, “Sonata: Self-supervised learning of reliable point representations,” inCVPR, 2025, pp. 22 193– 22 204

    2025

  24. [32]

    Tartanground: A large-scale dataset for ground robot perception and navigation,

    M. Patel, F. Yang, Y . Qiu, C. Cadena, S. Scherer, M. Hutter, and W. Wang, “Tartanground: A large-scale dataset for ground robot perception and navigation,”arXiv preprint arXiv:2505.10696, 2025

    2025 arXiv

  25. [33]

    Point transformer v3: Simpler faster stronger,

    X. Wu, L. Jiang, P.-S. Wang, Z. Liu, X. Liu, Y . Qiao, W. Ouyang, T. He, and H. Zhao, “Point transformer v3: Simpler faster stronger,” inCVPR, 2024, pp. 4840–4851

    2024

  26. [34]

    Loftup: Learning a coordinate-based feature upsampler for vision foundation models,

    H. Huang, A. Chen, V . Havrylov, A. Geiger, and D. Zhang, “Loftup: Learning a coordinate-based feature upsampler for vision foundation models,” inICCV, 2025, pp. 9913–9923

    2025

  27. [35]

    A survey on self-supervised learning: Algorithms, applications, and future trends,

    J. Gui, T. Chen, J. Zhang, Q. Cao, Z. Sun, H. Luo, and D. Tao, “A survey on self-supervised learning: Algorithms, applications, and future trends,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 9052–9071, 2024

    2024

  28. [36]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” inCVPR, 2020, pp. 9729–9738

    2020

  29. [37]

    Unsupervised learning of visual features by contrasting cluster assign- ments,

    M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin, “Unsupervised learning of visual features by contrasting cluster assign- ments,”Advances in neural information processing systems, vol. 33, pp. 9912–9924, 2020

    2020

  30. [38]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” inCVPR, 2022, pp. 16 000– 16 009

    2022

  31. [39]

    Emerging properties in self-supervised vision trans- formers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision trans- formers,” inICCV, 2021, pp. 9650–9660

    2021

  32. [40]

    ibot: Image bert pre-training with online tokenizer,

    J. Zhou, C. Wei, H. Wang, W. Shen, C. Xie, A. Yuille, and T. Kong, “ibot: Image bert pre-training with online tokenizer,”arXiv preprint arXiv:2111.07832, 2021

    2021 arXiv

  33. [41]

    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 ECCV. Springer, 2020, pp. 574–591

    2020

  34. [42]

    Transfusion: Robust lidar-camera fusion for 3d object detection with transformers,

    X. Bai, Z. Hu, X. Zhu, Q. Huang, Y . Chen, H. Fu, and C.-L. Tai, “Transfusion: Robust lidar-camera fusion for 3d object detection with transformers,” inCVPR, 2022, pp. 1090–1099

    2022

  35. [43]

    3dmv: Joint 3d-multi-view prediction for 3d semantic scene segmentation,

    A. Dai and M. Nießner, “3dmv: Joint 3d-multi-view prediction for 3d semantic scene segmentation,” inECCV, 2018, pp. 452–468

    2018

  36. [44]

    Bidirectional projection network for cross dimension scene understanding,

    W. Hu, H. Zhao, L. Jiang, J. Jia, and T.-T. Wong, “Bidirectional projection network for cross dimension scene understanding,” in CVPR, 2021, pp. 14 373–14 382

    2021

  37. [45]

    Lidar-camera panoptic segmentation via geometry-consistent and semantic-aware alignment,

    Z. Zhang, Z. Zhang, Q. Yu, R. Yi, Y . Xie, and L. Ma, “Lidar-camera panoptic segmentation via geometry-consistent and semantic-aware alignment,” inICCV, 2023, pp. 3662–3671

    2023

  38. [46]

    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

    2021 arXiv

  39. [47]

    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,” inCVPR, 2022, pp. 9891–9901

    2022

  40. [48]

    Three pillars improving vision foundation model distillation for lidar,

    G. Puy, S. Gidaris, A. Boulch, O. Sim ´eoni, C. Sautier, P. P ´erez, A. Bursuc, and R. Marlet, “Three pillars improving vision foundation model distillation for lidar,” inCVPR, 2024, pp. 21 519–21 529

    2024

  41. [49]

    Dino in the room: Leveraging 2d foundation models for 3d segmentation,

    K. A. Zeid, K. Yilmaz, D. de Geus, A. Hermans, D. Adrian, T. Linder, and B. Leibe, “Dino in the room: Leveraging 2d foundation models for 3d segmentation,”arXiv preprint arXiv:2503.18944, 2025

    2025

  42. [50]

    Concerto: Joint 2d-3d self-supervised learning emerges spatial rep- resentations,

    Y . Zhang, X. Wu, Y . Lao, C. Wang, Z. Tian, N. Wang, and H. Zhao, “Concerto: Joint 2d-3d self-supervised learning emerges spatial rep- resentations,”Advances in Neural Information Processing Systems, vol. 38, pp. 69 498–69 522, 2026

    2026

  43. [51]

    Sinkhorn distances: Lightspeed computation of optimal transport,

    M. Cuturi, “Sinkhorn distances: Lightspeed computation of optimal transport,”Advances in neural information processing systems, vol. 26, 2013

    2013

  44. [52]

    System design of the ultra mobility vehicle: A driving, bal- ancing, and jumping bicycle robot,

    B. Bokser, D. Gonzalez, A. Preston, A. Bahner, A. Wollschl ¨ager, A. Ilvonen, A. Eckert-Erdheim, A. Khadke, B. Hammoud, D. Molinaro et al., “System design of the ultra mobility vehicle: A driving, bal- ancing, and jumping bicycle robot,”arXiv preprint arXiv:2602.22118, 2026

    2026

  45. [53]

    The lov ´asz-softmax loss: A tractable surrogate for the optimization of the intersection-over- union measure in neural networks,

    M. Berman, A. R. Triki, and M. B. Blaschko, “The lov ´asz-softmax loss: A tractable surrogate for the optimization of the intersection-over- union measure in neural networks,” inCVPR, 2018, pp. 4413–4421

    2018

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.