Pith. sign in

REVIEW 3 major objections 6 minor 38 references

Long-Tailed 3D Point Cloud Dataset Distillation

T0 review · 3 major / 6 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read Long-tailed point cloud distillation works better when the synthetic budget and supervision follow the original imbalance instead of forcing balance.

desk verdict Real dual-imbalance gap in 3D distillation with solid gains; ASB’s benefit score is the softest link, not the whole story. read the letter →

arxiv 2607.26763 v1 pith:TX6XBCYL submitted 2026-07-29 cs.CV

Jiahao You , Xu Han , Jinfeng Xu , Xianzhi Li This is my paper
classification cs.CV
keywords datasetdistillation3Dpointcloudslong-tailedrecognitiondistributionmatchingsyntheticbudgetallocationdualimbalanceShapeNet55
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

Point cloud datasets are often long-tailed in both training and test splits, unlike the usual image setup of a skewed train set and a balanced test set. Existing point-cloud distillation methods ignore that fact and either chase geometry alone or build class-balanced synthetic sets. This paper argues that under a fixed total synthetic budget the right goal is to preserve the original long-tailed training utility: keep head classes well represented while still covering tails. It does so with Adaptive Synthetic Budgeting, which assigns more or fewer synthetic samples per class from class size and an estimated benefit score, then optimizes those samples with 3D Long-Tailed Distribution Matching—global plus local feature alignment plus prior-aware expert supervision that tightens constraints on tails and relaxes them on heads. On standard benchmarks the approach raises classification accuracy by as much as 7 points on ShapeNet55 versus strong baselines, and the same ideas transfer to part segmentation and to long-tailed image sets under weighted evaluation.

What carries the argument

Adaptive Synthetic Budgeting (ASB) plus 3D Long-Tailed Distribution Matching (3D-LTDM): ASB sets per-class synthetic counts from square-root-scaled class size and a benefit gap between a full expert and small balanced proxies; 3D-LTDM then optimizes the samples via Global-Local Feature Alignment and Prior-Aware Supervision (logit adjustment by log class prior).

What would settle it

Replace ASB’s benefit estimator with an oracle that knows true test frequencies (or with a clearly better benefit measure) and check whether the reported gains on ShapeNet55 and ModelNet40 shrink or vanish; if they do, the allocation rule—not the matching losses—was carrying the result.

Watch

Extended reading notes

Core claim

Under dual train–test imbalance and a fixed total synthetic budget, allocating class-wise budgets by quantity and sample benefit, then matching distributions with global-local feature alignment and class-prior-aware expert supervision, preserves the original long-tailed training utility better than class-balanced or geometry-only distillation, yielding large accuracy gains (up to 7.0 points on ShapeNet55).

Load-bearing premise

That a benefit score computed on a held-out slice of the training set—full expert accuracy minus accuracy of tiny balanced proxies—reliably tells how many synthetic samples each class should get under the true dual-imbalanced test distribution.

Editorial extensions

If this is right

  • Point-cloud distillation protocols should stop defaulting to equal samples per class when both train and test are long-tailed.
  • Under limited budget, preserving head-class mass while giving tails enough coverage beats forcing a balanced synthetic set.
  • Global class matching alone is insufficient when tail features are sparse; local cluster matching helps cover them.
  • Class-prior logit adjustment can keep tail synthetics recognizable without collapsing head-class diversity.
  • The same budget-then-match pattern extends to part segmentation and to long-tailed images under frequency-weighted accuracy.

Reading between the lines

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

  • If train and test imbalance factors diverge sharply, ASB’s train-only benefit scores may systematically over- or under-fund classes that matter at test time—suggesting a natural next experiment with controlled train–test IF mismatch.
  • The dual-imbalance framing may matter for other 3D tasks (detection, scene understanding) where object frequencies are naturally skewed on both splits.
  • Geometry-aware point-cloud distillers that ignore class counts may keep underperforming simple image-style matchers until they add an explicit distribution term.
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 / 6 minor

Summary. The paper studies dataset distillation for 3D point clouds under dual long-tailed imbalance (both train and test). It argues that class-balanced synthetic sets are inappropriate when the test distribution is also skewed, and proposes a budget-then-optimize pipeline: Adaptive Synthetic Budgeting (ASB) allocates a fixed total budget M using class size and a sample-benefit score from expert vs. small balanced proxies (Eqs. 3–5), then 3D-LTDM optimizes the synthetics via Global-Local Feature Alignment (class- and cluster-level M3D matching) and Prior-Aware Supervision (logit adjustment by class prior, Eqs. 12–14). Experiments on ModelNet10/40, ShapeNet55, and ScanObjectNN report consistent gains over coreset, image DD, long-tailed DD, and point-cloud DD baselines (up to +7.0 OA on ShapeNet55 at PPC=10), with supporting results on part segmentation, cross-architecture transfer, ablations, head/middle/tail breakdown, initialization robustness, efficiency, and a weighted-accuracy image transfer setting.

Significance. If the results hold, this is a clear and timely contribution: it identifies dual imbalance as a structural difference between standard 3D benchmarks and conventional image long-tailed DD, and shows that geometry-only point-cloud distillation is insufficient under that regime. The empirical package is relatively strong for the area (multiple datasets, five-run means±std, segmentation, cross-backbone transfer, efficiency). Framing the objective as preserving original long-tailed utility rather than forcing balance is a useful conceptual correction for the community. The main novelty is the problem setting plus the coupled ASB + long-tail-aware DM design, not a new matching primitive alone.

major comments (3)
  1. [Methodology (ASB); Table 5] Methodology, Adaptive Synthetic Budgeting (Eqs. 3–5) and Table 5: ASB’s claimed benefit signal gc = max(a^exp_c − ā_c, 0) + ε is not isolated. The only ASB ablation removes the whole module (uniform mc), which confounds quantity scaling √nc with the benefit term. A load-bearing check for the paper’s attribution is missing: freeze gc ≡ 1 (pure √nc / quantity allocation) versus full (gc √nc)^α, and optionally report rank correlation of gc with held-out marginal OA gain when adding one synthetic slot per class. Without this, the large ShapeNet55 lift cannot be confidently credited to “expected benefit” rather than head-leaning reallocation plus 3D-LTDM. Please add this ablation (and, if space, sensitivity of α).
  2. [Table 6; Experiments / Main Results] Table 6 and the dual-imbalance objective (Introduction; Eq. 2): head/middle accuracy rises while tail is flat or lower (e.g., MN40 tail 70.4→66.0). That is consistent with preserving long-tailed test mass, but the paper should state the evaluation protocol more sharply—report both overall accuracy and macro-averaged accuracy (or IF-matched weighted metrics) on all main tables, not only OA—so readers can separate “matches the skewed test prior” from “improves recognition of rare classes.” Otherwise the +7.0 headline can be read as prior-matching rather than better distillation of hard classes.
  3. [Table 1; Related Work] Related Work and baselines: image long-tailed DD methods (LTDD, and RLDD only on CIFAR) are designed for balanced test sets; applying them unchanged is a reasonable baseline, but the comparison would be fairer if at least one strong point-cloud DD method were also run under the same non-uniform budget that ASB produces (fixed {mc}, optimize with their loss). That would separate “better allocation” from “better matching under a given allocation,” which the current Table 1 mixes.
minor comments (6)
  1. [Experimental Setup] Hyperparameters α, τ, λ_local, λ_pas, m_min, ε are listed in spirit but not given default values or a selection protocol in the provided text; add a short implementation paragraph or appendix table for reproducibility.
  2. [Problem Formulation] Eq. (2) uses an absolute expected loss gap to the full-data model; in practice optimization uses LLTDM. Briefly clarify that Eq. (2) is the ideal objective and not the training loss.
  3. [Figure 1] Figure 1 is helpful; state explicitly how train/test IF values were computed for each 3D dataset (official splits, class counts) so the dual-imbalance claim is auditable.
  4. [Abstract / Introduction] Typos/spacing artifacts from PDF extraction appear throughout (e.g., “First,wedesign”, “3D-LTDM tooptimize”); clean compounding in the camera-ready.
  5. [Global-Local Feature Alignment] K_c = m_c ties local clusters to the synthetic count; a one-sentence justification or small sensitivity check would help when m_c is very small (PPC=1).
  6. [Table 4] Table 4 uses weighted accuracy to simulate dual imbalance on CIFAR-LT; note clearly that this is a proxy, not a true long-tailed test set, to avoid over-claiming image-domain dual imbalance.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical method with external baselines; objectives do not force reported accuracies by construction.

full rationale

This is a standard methods-and-benchmarks paper. ASB (Eqs. 3–5) defines a heuristic budget allocation from held-out training-split expert/proxy accuracies and class sizes; 3D-LTDM (GLFA + PAS, Eqs. 9–15) is an optimization objective over synthetic point clouds. Neither quantity is algebraically identical to the claimed test OA/mIoU gains. Evaluation is against external coreset and distillation baselines (Table 1–4), the full original dataset, cross-architecture transfer, and ablations (Table 5). PAS’s π_c is the ordinary empirical class prior from training counts—standard long-tail practice, not a fitted parameter renamed as a prediction. No load-bearing self-citation uniqueness theorem, no ansatz smuggled as external fact, and no self-definitional loop (X defined as Y then “derived” as Y). Concerns that gc is unvalidated or that √nc reweighting may drive gains are correctness/ablation issues, not circularity. Score 0; steps empty.

Assumptions & free parameters 6 free parameters · 6 assumptions · 2 invented entities

The claim rests on standard distribution-matching distillation practice, empirical long-tail priors, and several hand-chosen allocation/loss knobs. No physical entities are postulated; the new content is algorithmic modules whose value is only evidenced by the paper’s own benchmarks.

free parameters (6)
  • α (ASB allocation sharpness)
    Exponent on (g_c √n_c) in Eq. 4 controlling how aggressively budgets follow benefit×size; chosen by authors, not derived.
  • τ (logit-adjustment strength)
    Scales τ log π_c in Prior-Aware Supervision (Eq. 13); strength of tail vs head supervision is a free knob.
  • λ_local, λ_pas, λ_cls (loss weights)
    Trade off global/local matching and classification terms (Eqs. 6, 11, 15); standard fitted hyperparameters.
  • m_min (per-class minimum synthetic count)
    Floor in Eq. 5 guaranteeing tail coverage before weighted remainder allocation.
  • ϵ (benefit-score smoother)
    Additive constant in g_c = max(a_exp_c − ā_c, 0) + ϵ (Eq. 3).
  • K_c = m_c (clusters per class) = K_c equals class synthetic budget m_c
    Local alignment sets one K-means cluster per synthetic slot; couples budget to clustering granularity by design choice.
assumptions (6)
  • domain assumption Distribution matching on intermediate features plus an expert classification term is a valid surrogate for retaining training utility of the full set (standard DM backbone).
    Invoked when moving from problem (2) to L_DM and then L_LTDM in §3D Long-Tailed Distribution Matching.
  • ad hoc to paper When the test distribution is long-tailed, the synthetic set should preserve original imbalanced utility rather than enforce class balance under fixed M.
    Core design premise in Introduction contrasting image LT distillation; if false, balanced baselines could be preferable.
  • ad hoc to paper Sample-benefit gap between full expert and small balanced proxy models on a train-derived validation split predicts marginal value of extra synthetic samples per class.
    Defines g_c in Adaptive Synthetic Budgeting (Eq. 3); load-bearing for ASB.
  • domain assumption Empirical training class frequencies π_c are appropriate priors for logit adjustment during synthetic optimization.
    PAS Eqs. 12–14; standard long-tail classification device applied to distillation.
  • domain assumption K-means clusters on expert features capture stable intra-class modes useful as local matching targets, with fixed assignments during optimization.
    GLFA local term (Eq. 10) and initialization around centroids.
  • ad hoc to paper Square-root class-size scaling sufficiently prevents head-class budget domination while still reflecting quantity.
    w_c = (g_c √n_c)^α in Eq. 4; modeling choice without uniqueness proof.
invented entities (2)
  • Adaptive Synthetic Budgeting (ASB)
    purpose: Allocate limited synthetic slots across long-tailed classes using benefit scores and √n_c weights.
    Named module introduced here; only validated inside this paper’s benchmarks.
  • 3D Long-Tailed Distribution Matching (GLFA + PAS)
    purpose: Optimize synthetic point clouds under class imbalance via global-local feature alignment and prior-adjusted expert CE.
    Composite objective L_LTDM specific to this work; no external replication cited.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Long-Tailed 3D Point Cloud Dataset Distillation." pith.science (2026). https://pith.science/paper/TX6XBCYL

@misc{pith2026260726763,
  author       = {Pith},
  title        = {Pith review of: Long-Tailed 3D Point Cloud Dataset Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TX6XBCYL}},
  note         = {Machine review of arXiv:2607.26763}
}
read the original abstract

Dataset distillation compresses large-scale datasets into compact synthetic sets while preserving their training utility, enabling efficient 3D point cloud training. Current point cloud dataset distillation methods only tackle geometric and representation challenges while ignoring the distributional imbalance prevalent in point cloud datasets where both training and test splits follow long-tailed class distributions. To our knowledge, we present the first study on long-tailed point cloud dataset distillation. Rather than focusing primarily on geometric and representation properties or simply constructing a class-balanced synthetic set, our framework explicitly accounts for long-tailed class distributions via two core modules. First, we design Adaptive Synthetic Budgeting to allocate class-wise synthetic budgets according to class quantity and the expected benefit of additional synthetic samples. Given the allocated budgets, we further design 3D Long-Tailed Distribution Matching to optimize synthetic point clouds through Global-Local Feature Alignment and Prior-Aware Supervision. The former preserves both global class distributions and diverse intra-class structures, while the latter provides class-dependent expert supervision to keep tail-class samples recognizable while maintaining diverse head-class patterns. Extensive experiments demonstrate the effectiveness of our method, lifting classification accuracy by 7.0 points on ShapeNet55 against state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2607.26763 by the authors.

Figure 1
Figure 1. Train–test class imbalance across benchmarks. The [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our framework. Given an original long-tailed dataset and a fixed total synthetic budget, ASB adaptively [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. PCA visualization of feature-space distributions on [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Feature-space visualization with PCA for original [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Accuracy curves under different initialization [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 1 linked inside Pith

  1. [1]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    3d shapenets: A deep representation for volumetric shapes , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  2. [2]

    Chang, Angel X. and Funkhouser, Thomas and Guibas, Leonidas and Hanrahan, Pat and Huang, Qixing and Li, Zimo and Savarese, Silvio and Savva, Manolis and Song, Shuran and Su, Hao and Xiao, Jianxiong and Yi, Li and Yu, Fisher , number =

  3. [3]

    International Conference on Computer Vision (ICCV) , year =

    Revisiting Point Cloud Classification: A New Benchmark Dataset and Classification Model on Real-World Data , author =. International Conference on Computer Vision (ICCV) , year =

  4. [4]

    Proceedings of the International Conference on Machine Learning (ICML) , year=

    Point Cloud Dataset Distillation , author=. Proceedings of the International Conference on Machine Learning (ICML) , year=

  5. [5]

    Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , year=

    Dataset Distillation of 3D Point Clouds via Distribution Matching , author=. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , year=

  6. [6]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages =

    Zhang, Wenxiao and Wang, Ziqi and Xu, Li and Yang, Xun and Liu, Jun , title =. Proceedings of the 32nd ACM International Conference on Multimedia , pages =. 2024 , isbn =. doi:10.1145/3664647.3680767 , abstract =

    arXiv 2024

  7. [7]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Pointnet: Deep learning on point sets for 3d classification and segmentation , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  8. [8]

    Advances in Neural Information Processing Systems , pages=

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space , author=. Advances in Neural Information Processing Systems , pages=

Show all 38 references
  1. [9]

    and Hu, Shi-Min , year=

    Guo, Meng-Hao and Cai, Jun-Xiong and Liu, Zheng-Ning and Mu, Tai-Jiang and Martin, Ralph R. and Hu, Shi-Min , year=. PCT: Point cloud transformer , volume=. Computational Visual Media , publisher=. doi:10.1007/s41095-021-0229-5 , number=

    doi

  2. [10]

    ACM Transactions on Graphics (TOG) , year=

    Dynamic Graph CNN for Learning on Point Clouds , author=. ACM Transactions on Graphics (TOG) , year=

  3. [11]

    Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual

    Xu Ma and Can Qin and Haoxuan You and Haoxi Ran and Yun Fu , booktitle=. Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual. 2022 , url=

    2022

  4. [12]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Wu, Wenxuan and Qi, Zhongang and Fuxin, Li , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

  5. [13]

    Proceedings of the 26th Annual International Conference on Machine Learning , pages =

    Welling, Max , title =. Proceedings of the 26th Annual International Conference on Machine Learning , pages =. 2009 , isbn =. doi:10.1145/1553374.1553517 , abstract =

    2009

  6. [14]

    International Conference on Learning Representations , year=

    Active Learning for Convolutional Neural Networks: A Core-Set Approach , author=. International Conference on Learning Representations , year=

  7. [15]

    Proceedings of the International Conference on Learning Representations (ICLR) , year=

    Dataset Condensation with Gradient Matching , author=. Proceedings of the International Conference on Learning Representations (ICLR) , year=

  8. [16]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    Dataset Condensation with Distribution Matching , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

  9. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Dataset Distillation by Matching Training Trajectories , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  10. [18]

    Zhang, Hansong and Li, Shikun and Lin, Fanzhao and Wang, Weiping and Qian, Zhenxing and Ge, Shiming , booktitle=

  11. [19]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Distilling Long-tailed Datasets , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  12. [20]

    Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , year=

    Rectifying Soft-Label Entangled Bias in Long-Tailed Dataset Distillation , author=. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , year=

  13. [21]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year=

    Rethinking Long-tailed Dataset Distillation: A Uni-Level Framework with Unbiased Recovery and Relabeling , author=. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year=

  14. [22]

    arXiv preprint arXiv:2505.13300 , year=

    DD-Ranking: Rethinking the Evaluation of Dataset Distillation , author=. arXiv preprint arXiv:2505.13300 , year=

  15. [23]

    2024 , booktitle=

    M3D: Dataset Condensation by Minimizing Maximum Mean Discrepancy , author=. 2024 , booktitle=

    2024

  16. [24]

    Proceedings of the International Conference on Learning Representations (ICLR) , year=

    Parameterization-Based Dataset Distillation of 3D Point Clouds through Learnable Shape Morphing , author=. Proceedings of the International Conference on Learning Representations (ICLR) , year=

  17. [25]

    arXiv preprint arXiv:1811.10959 , year=

    Dataset Distillation , author=. arXiv preprint arXiv:1811.10959 , year=

    arXiv

  18. [26]

    Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , year=

    Squeeze, Recover and Relabel: Dataset Condensation at ImageNet Scale From A New Perspective , author=. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , year=

  19. [27]

    2009 , institution=

    Learning multiple layers of features from tiny images , author=. 2009 , institution=

    2009

  20. [28]

    Ran, Fengli and Pu, Xiao and Liu, Bo and Bi, Xiuli and Xiao, Bin , booktitle=

  21. [29]

    CVPR , year=

    Class-Balanced Loss Based on Effective Number of Samples , author=. CVPR , year=

  22. [30]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    BCE3S: Binary Cross-Entropy Based Tripartite Synergistic Learning for Long-tailed Recognition , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  23. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Improved distribution matching for dataset condensation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  24. [32]

    Wang, Kai and Zhao, Bo and Peng, Xiangyu and Zhu, Zheng and Yang, Shuo and Wang, Shuo and Huang, Guan and Bilen, Hakan and Wang, Xinchao and You, Yang , booktitle=

  25. [33]

    Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms , pages =

    K-means++: The Advantages of Careful Seeding , author =. Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms , pages =. 2007 , publisher =

    2007

  26. [34]

    Proceedings of the International Conference on Learning Representations (ICLR) , year=

    Towards Lossless Dataset Distillation via Difficulty-Aligned Trajectory Matching , author=. Proceedings of the International Conference on Learning Representations (ICLR) , year=

  27. [35]

    Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , year=

    Elucidating the Design Space of Dataset Condensation , author=. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , year=

  28. [36]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Efficient Dataset Distillation via Minimax Diffusion , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  29. [37]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year=

    Post Training Quantization for Efficient Dataset Condensation , author=. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year=

  30. [38]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Large-Scale Long-Tailed Recognition in an Open World , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.