Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

A Large-Scale Study on Video Action Dataset Condensation

T0 review · 4 major / 6 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 A large-scale study on video dataset condensation compares sample selection and dataset distillation, finding that distillation wins on large datasets and selection on small ones.

desk verdict Useful benchmark study whose headline claim about distillation winning on hard datasets is undercut by a top-1/top-5 metric mismatch and paper-thin margins. read the letter →

arxiv 2412.21197 v2 pith:5KGFZ54B submitted 2024-12-30 cs.CV

Yang Chen , Sheng Guo , Bo Zheng , Limin Wang This is my paper · ORCID
classification cs.CV
keywords videodatasetcondensationactionrecognitiondistillationsampleselectionevaluationprotocoltemporalprocessingsliding-windowsampling
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 conducts a systematic, large-scale empirical comparison of dataset condensation methods for video action recognition, adapting three image-domain algorithms to the spatio-temporal domain. It claims that the best method class depends on dataset difficulty: sample selection (RDED) is superior on small, easier datasets like HMDB51 and UCF101, while dataset distillation (DATM) wins on large, more challenging datasets like SSv2 and K400. The study also proposes a unified evaluation protocol to make such comparisons fair and introduces sliding-window sampling for temporal processing. If correct, this provides concrete guidance on which condensation approach to use at different data scales.

What carries the argument

The central machinery is the empirical comparison framework: three condensation algorithms (DATM for trajectory matching, EDC for statistical matching, and RDED for score-based sample selection) extended to video with sliding-window sampling, combined with a unified evaluation protocol that fixes the labeling method (Multi-SL), loss function (MSE-GT), augmentation (resized crop and horizontal flip), training length (300 epochs), and batch size (proportional to IPC). Sliding-window sampling sequentially samples overlapping clips along time to maintain temporal coherence, and the protocol standardizes downstream evaluation so that method comparisons are not artifacts of evaluation settings.

What would settle it

An experiment that would settle it is to repeat the SSv2 and K400 comparisons under the alternative settings the paper itself found can flip conclusions, such as hard labeling, CutMix augmentation, or KL loss; if DATM no longer outperforms RDED there, the claim that distillation wins on challenging scenarios collapses.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the relative performance of sample selection versus dataset distillation flips with dataset scale and diversity. Under a unified evaluation protocol using Multi-SL labeling, MSE-GT loss, and no CutMix, RDED (sample selection) achieves better top-1 accuracy than DATM on HMDB51 and UCF101, while DATM (trajectory-matching distillation) achieves the best top-5 accuracy on SSv2 and K400. The paper attributes this to the larger scale and greater diversity of the latter two datasets, arguing that distillation methods preserve information more effectively in challenging scenarios.

Load-bearing premise

The unified evaluation protocol (Multi-SL labeling, MSE-GT loss, no CutMix, 300 epochs) provides a neutral comparison; if those settings favor one method class, the main result is a protocol artifact.

Editorial extensions

If this is right

  • On large video datasets, dataset distillation methods such as DATM are the better choice; on small datasets, sample selection such as RDED is sufficient and more efficient.
  • The proposed unified evaluation protocol provides a common yardstick for future video dataset condensation research to report results.
  • Sliding-window sampling should replace segment sampling for temporal processing, since it preserves temporal coherence and improves accuracy.
  • When allocating a fixed condensation budget, the number of videos (IPC) matters more than the number of frames per video.
  • Labeling methods can affect results more than the condensation algorithm itself, so they must be fixed when comparing methods.

Reading between the lines

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

  • The 'challenging scenario' explanation is correlational; a direct test would manipulate dataset diversity at fixed size to see whether DATM's advantage tracks task difficulty.
  • The protocol's settings are tuned on UCF101 with IPC=1, and the paper itself shows that settings like CutMix or KL loss can flip conclusions, so the fairness of the yardstick is uncertain outside those settings.
  • A trainable interpolator, which the paper does not provide, could close the gap between distillation and selection by better condensing temporal information.
  • The findings suggest that trajectory matching methods merit further development to reduce their high memory cost, since they are the winning class on large datasets.
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 / 6 minor

Summary. This paper presents a large-scale empirical study of video dataset condensation. It adapts three image-domain condensation algorithms (RDED as sample selection; DATM and EDC as dataset distillation) to video by introducing temporal sampling and interpolation schemes, including a proposed sliding-window sampling. It systematically ablates evaluation settings (labeling methods, data augmentation, loss functions) on UCF101 and derives a unified evaluation protocol. Main experiments on HMDB51, UCF101, SSv2, and K400 report that RDED is best on the smaller/easier datasets while DATM is best on the larger/more challenging ones, and the paper concludes that dataset distillation methods surpass sample selection in challenging scenarios while selection methods excel in easier ones. The paper also reports that labeling methods strongly affect performance and that sliding-window sampling improves over segment sampling.

Significance. If the central observation held, the paper would be a valuable reference for choosing condensation strategies in video and for benchmarking future methods. The strengths are the breadth of the study (four datasets, several backbones, cross-architecture evaluation), the systematic ablations, the release of code, and the explicit statement of limitations. However, the cross-dataset ranking that grounds the main conclusion is confounded by inconsistent evaluation metrics and lacks statistical support; the paper is best viewed as a comprehensive empirical contribution whose headline claim requires further evidence.

major comments (4)
  1. [Section 4.4, Table 8] The central observation that DATM outperforms RDED on SSv2 and K400 is based on top-5 accuracy for those datasets and top-1 accuracy for HMDB51/UCF101. Top-5 and top-1 have different chance levels and sensitivities, so the relative ranking of the two methods across dataset pairs is not directly comparable. Since observation (iii) in the abstract and Section 4.4 explicitly contrasts the two pairs, the metric mismatch is load-bearing for the paper's main conclusion. Please report a common metric (at least top-1) for all datasets, and additionally report top-5 for completeness.
  2. [Section 4.4, Table 8] The claimed advantage of DATM over RDED on SSv2 and K400 is tiny: at IPC=1 the differences are 0.2 points (SSv2 12.9 vs 12.7; K400 12.1 vs 11.9) and at IPC=5 the K400 difference is also 0.2 points (16.7 vs 16.5). No error bars or multiple-seed results are reported, so these differences are within the noise one would expect from a single run of the evaluation network. To support the ranking, please report mean and standard deviation over at least three seeds for the key comparisons in Table 8, or state the precision of the numbers.
  3. [Section 4.1, Tables 2 and 3] The unified evaluation protocol is selected post-hoc from UCF101 IPC=1 ablations (MSE-GT over KL, no CutMix, Multi-SL), yet Table 8—the table supporting the main cross-dataset claim—does not use the proposed protocol but uses soft labeling for comparability with prior work. Since the paper itself shows that evaluation settings can flip individual comparisons (e.g., DATM without CutMix at 12.1 vs. RDED with CutMix at 11.0 under soft labels), the neutrality of the chosen protocol on SSv2/K400 cannot be assumed. Please provide protocol ablations (labeling, loss, CutMix) on at least one of the two large datasets, or justify transferability from UCF101.
  4. [Section 4.4, observation (iii)] The distinction between 'challenging' and 'easier' datasets is confounded with several variables that change between the two pairs: evaluation metric (top-1 vs top-5), input resolution (112×112 vs 56×56), and condensation ratio (e.g., 0.4‰ vs 0.05‰ for UCF101 vs K400 at IPC=1). These confounds make the causal attribution to 'challenging scenarios' uncertain. Please either soften the claim to a description of the observed rankings or conduct a controlled comparison, e.g., evaluating all methods on the same metric and resolution, or matching the condensation ratio across datasets.
minor comments (6)
  1. [Section 4.1] Typo: 'using MES-GT loss' should be 'MSE-GT loss'.
  2. [Section 4.1] Grammar: 'we adopted a evaluation protocol' should be 'we adopted an evaluation protocol'.
  3. [Appendix B] The batch size in Eq. (5) uses a dataset-specific base (10/20/40) without justification; please comment on the sensitivity of the results to this choice.
  4. [Section 4.2, Table 4] The drop from interpolation differs sharply across methods (e.g., DATM from 8.5 to 5.5, while EDC from 12.0 to 11.8); the text says interpolation does not work for all methods but could be clearer on this interaction.
  5. [Table 8] The rows for DM, MTT, and FRePo are missing on several datasets; consider marking unavailable entries explicitly rather than leaving blanks.
  6. [Section 4.4] The sentence 'For sample selection methods, our results show that Random outperforms Herding in most cases. This is violated with the intuition.' should be rephrased ('This goes against intuition').

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the paper's conclusions are empirical benchmark measurements under a stated protocol; at most one non-load-bearing self-citation in related work.

full rationale

The paper's central observations—labeling methods dominate performance, sliding-window sampling helps, and RDED versus DATM rankings differ across datasets—are empirical measurements on HMDB51, UCF101, SSv2, and K400, with training settings stated. The unified evaluation protocol (Multi-SL labeling, MSE-GT loss, no CutMix, batch size proportional to IPC) is a standardization choice justified by ablations; it is not a fitted parameter renamed as a prediction, and the method comparisons are not mathematically forced by the protocol's definition. The 'challenging versus easy' distinction rests on independent dataset properties such as scale, category count, and diversity, not on the experimental outcomes. Citations are used to import existing algorithms or backbones, not to justify the paper's conclusions by self-reference. The only self-citation is [37] (VideoMAE) in the action-recognition background, which is not load-bearing for any condensation claim. Concerns about top-1 versus top-5 metric comparability or protocol selection are correctness/generality risks, not circularity. Accordingly, no step in the derivation reduces to its own input by construction.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper is an empirical benchmark; it introduces no new theoretical entities. Its central claims rest on the choice of evaluation protocol (labeling, loss, augmentation, batch size, epochs), which was selected after inspecting ablations, and on the assumption that MiniC3D is a representative evaluation architecture. The DATM hyperparameters are fitted to UCF101. No ad hoc entities are invented.

free parameters (4)
  • Evaluation protocol: base batch size (10/20/40 per dataset) = 10 (HMDB51/UCF101), 20 (SSv2), 40 (K400)
    Chosen to keep iteration count fixed across datasets and IPC; directly affects training dynamics of the evaluation network and therefore all reported numbers.
  • Evaluation protocol: loss function (MSE-GT vs KL) = MSE-GT
    Selected after ablation on UCF101 IPC=1; RDED actually scores higher with KL (15.0 vs 14.3), so the choice is not neutral for all methods.
  • Evaluation protocol: labeling method (Multi-SL vs soft vs hard) = Multi-SL for ablations; soft in Table 8
    Labeling choice changes accuracy by up to 8.6%; the protocol uses Multi-SL except when comparing to prior work.
  • DATM hyperparameters (mapping scope, syn-step, batch-syn) = Grid-searched on UCF101 IPC=1, then adjusted per dataset
    Hyperparameters for the distillation method are fitted on one dataset and transfer; this is inherited from the base method but affects results.
assumptions (3)
  • domain assumption MiniC3D is a sufficient evaluation architecture for comparing condensation methods
    All main rankings are measured with MiniC3D as the evaluation network; cross-architecture validation is only on UCF101 IPC=1, so conclusions on SSv2/K400 may not transfer to other backbones.
  • domain assumption Accuracy on the real validation set after training on the condensed set is a valid proxy for condensation quality
    Standard in the dataset condensation literature; assumed throughout the paper, not tested against other metrics like data efficiency or privacy.
  • domain assumption The condensation ratio (Nc*Tc)/(N*Tm) and IPC are meaningful compression metrics
    Used to define the study setup in Sec 3.1; the paper does not justify why these are the right measures of compression.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Large-Scale Study on Video Action Dataset Condensation." pith.science (2026). https://pith.science/paper/5KGFZ54B

@misc{pith2026241221197,
  author       = {Pith},
  title        = {Pith review of: A Large-Scale Study on Video Action Dataset Condensation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5KGFZ54B}},
  note         = {Machine review of arXiv:2412.21197}
}
read the original abstract

Recently, dataset condensation has made significant progress in the image domain. Unlike images, videos possess an additional temporal dimension, which harbors considerable redundant information, making condensation even more crucial. However, video dataset condensation still remains an underexplored area. We aim to bridge this gap by providing a large-scale study with systematic design and fair comparison. Specifically, our work delves into three key aspects to provide valuable empirical insights: (1) temporal processing of video data, (2) the evaluation protocol for video dataset condensation, and (3) adaptation of condensation algorithms to the space-time domain. From this study, we derive several intriguing observations: (i) labeling methods greatly influence condensation performance, (ii) simple sliding-window sampling is effective for temporal processing, and (iii) dataset distillation methods perform better in challenging scenarios, while sample selection methods excel in easier ones. Furthermore, we propose a unified evaluation protocol for the fair comparison of different condensation algorithms and achieve state-of-the-art results on four widely-used action recognition datasets: HMDB51, UCF101, SSv2 and K400. Our code is available at https://github.com/MCG-NJU/Video-DC.

Figures

Figures reproduced from arXiv: 2412.21197 by the authors.

Figure 1
Figure 1. Pipeline of our study. Our study includes three core el [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Temporal processing encompasses sampling and inter [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Conceptual visualization of three dataset condensation frameworks applied to video. Trajectory Matching (a) and Distribution [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance curve of various condensation algorithms [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Visualization of RDED for UCF101 IPC=1 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Visualization of EDC for UCF101 IPC=1 10 [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Temporal Saliency-Guided Distillation: A Scalable Framework for Distilling Video Datasets

    cs.CV 2025-05 conditional novelty 6.0 of 10

    Temporal saliency masks computed from inter-frame differences guide gradient updates and augmentation in a uni-level video dataset distillation framework, achieving state-of-the-art results on MiniUCF, HMDB51, Kinetic...

  2. The Evolution of Dataset Distillation: Toward Scalable and Generalizable Solutions

    cs.CV 2025-02 conditional novelty 3.0 of 10

    A 2023-2025 survey of dataset distillation that organizes matching, generative, decoupling, and selective methods and tabulates ImageNet-scale accuracy comparisons.

Reference graph

Works this paper leans on

55 extracted references · 40 canonical work pages · cited by 2 Pith papers

  1. [1]

    Is space-time attention all you need for video understanding? In ICML, page 4, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 3

    2021

  2. [2]

    Quo vadis, action recognition? a new model and the kinetics dataset

    Jo ˜ao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 4724–4733, 2017. 2, 5

    work page 2017

  3. [3]

    Dataset distillation by matching training trajectories

    George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In CVPR, 2022. 1, 2

    work page 2022

  4. [4]

    Efros, and Jun-Yan Zhu

    George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A. Efros, and Jun-Yan Zhu. Generalizing dataset distillation via deep generative prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3739–3748, 2023. 8

    work page 2023

  5. [5]

    Super-samples from kernel herding

    Yutian Chen, Max Welling, and Alex Smola. Super-samples from kernel herding. In Proceedings of the Twenty-Sixth Conference on Uncertainty in Artificial Intelligence , pages 109–116, 2010. 2

    work page 2010

  6. [6]

    DC- BENCH: Dataset condensation benchmark

    Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. DC- BENCH: Dataset condensation benchmark. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022. 8

    work page 2022

  7. [7]

    Scaling up dataset distillation to imagenet-1k with constant memory

    Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation to imagenet-1k with constant memory. In ICML, pages 6565–6590, 2023. 1, 2, 5

    work page 2023

  8. [8]

    Remember the past: Distilling datasets into addressable memories for neural net- works

    Zhiwei Deng and Olga Russakovsky. Remember the past: Distilling datasets into addressable memories for neural net- works. In NeurIPS, 2022. 1

    work page 2022

Show all 55 references
  1. [9]

    CondTSF: One-line plugin of dataset condensation for time series forecasting

    Jianrong Ding, Zhanyu Liu, Guanjie Zheng, Haiming Jin, and Linghe Kong. CondTSF: One-line plugin of dataset condensation for time series forecasting. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2024. 2

    2024

  2. [10]

    Slowfast networks for video recognition

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202–6211, 2019. 2

    2019

  3. [11]

    The ”something something” video database for learning and evaluating visual common sense,

    Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michal- ski, Joanna Materzy ´nska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The ”something something” video d...

  4. [12]

    Deepcore: A comprehensive library for coreset selection in deep learning,

    Chengcheng Guo, Bo Zhao, and Yanbing Bai. Deepcore: A comprehensive library for coreset selection in deep learning,

  5. [13]

    Towards lossless dataset distillation via difficulty-aligned trajectory matching

    Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. In The Twelfth In- ternational Conference on Learning Representations , 2024. 1, 2, 3, 5, 6

    2024

  6. [14]

    Characterizing structural regularities of labeled data in overparameterized models

    Ziheng Jiang, Chiyuan Zhang, Kunal Talwar, and Michael C Mozer. Characterizing structural regularities of labeled data in overparameterized models. In International Conference on Machine Learning, 2021. 2

    2021

  7. [15]

    Graph condensation for graph neural networks

    Wei Jin, Lingxiao Zhao, Shichang Zhang, Yozen Liu, Jiliang Tang, and Neil Shah. Graph condensation for graph neural networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2022. 2

    2022

  8. [16]

    Grad-match: Gradient matching based data subset selection for efficient deep model training

    Krishnateja Killamsetty, Sivasubramanian Durga, Ganesh Ramakrishnan, Abir De, and Rishabh Iyer. Grad-match: Gradient matching based data subset selection for efficient deep model training. In International Conference on Ma- chine Learning, 2021. 2

    2021

  9. [17]

    Glister: Generalization based data subset selection for efficient and robust learning

    Krishnateja Killamsetty, Durga Sivasubramanian, Ganesh Ramakrishnan, and Rishabh Iyer. Glister: Generalization based data subset selection for efficient and robust learning. In Association for the Advancement of Artificial Intelligence,

  10. [18]

    Dataset condensation via efficient synthetic- data parameterization

    Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic- data parameterization. In ICML, 2022. 1

    2022

  11. [19]

    Kuehne, H

    H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. Hmdb: A large video database for human motion recogni- tion. In ICCV, pages 2556–2563, 2011. 2, 5

    2011

  12. [20]

    Selmatch: Effectively scaling up dataset distillation via selection-based initializa- tion and partial updates by trajectory matching

    Yongmin Lee and Hye Won Chung. Selmatch: Effectively scaling up dataset distillation via selection-based initializa- tion and partial updates by trajectory matching. In Forty-first International Conference on Machine Learning, 2024. 9

    2024

  13. [21]

    Dataset distillation via factorization

    Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xinchao Wang. Dataset distillation via factorization. In NeurIPS, 2022. 1

    2022

  14. [22]

    Dataset condensation for time series classification via dual domain matching

    Zhanyu Liu, Ke Hao, Guanjie Zheng, and Yanwei Yu. Dataset condensation for time series classification via dual domain matching. In Proceedings of the ACM SIGKDD Con- ference on Knowledge Discovery and Data Mining (KDD) ,

  15. [23]

    Efficient dataset distillation using random feature ap- proximation

    Noel Loo, Ramin Hasani, Alexander Amini, and Daniela Rus. Efficient dataset distillation using random feature ap- proximation. In NeurIPS, 2022. 1, 2

    2022

  16. [24]

    Dataset distillation fixes dataset reconstruction attacks

    Noel Loo, Ramin Hasani, Mathias Lechner, Alexander Amini, and Daniela Rus. Dataset distillation fixes dataset reconstruction attacks. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 1

    2024

  17. [25]

    UniDetox: Universal detoxification of large lan- guage models via dataset distillation

    Huimin Lu, Masaru Isonuma, Junichiro Mori, and Ichiro Sakata. UniDetox: Universal detoxification of large lan- guage models via dataset distillation. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. 1, 2 11

    2025

  18. [26]

    Coresets for data-efficient training of machine learning mod- els

    Baharan Mirzasoleiman, Jeff Bilmes, and Jure Leskovec. Coresets for data-efficient training of machine learning mod- els. In International Conference on Machine Learning, 2020. 2

    2020

  19. [27]

    Dataset meta-learning from kernel ridge-regression

    Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. arXiv preprint arXiv:2011.00050, 2020. 1, 2

    2011 arXiv

  20. [28]

    Dataset distillation with infinitely wide convolutional networks

    Timothy Nguyen, Roman Novak, Lechao Xiao, and Jaehoon Lee. Dataset distillation with infinitely wide convolutional networks. In NeurIPS, 2021. 1, 2

    2021

  21. [29]

    A la- bel is worth a thousand images in dataset distillation

    Tian Qin, Zhiwei Deng, and David Alvarez-Melis. A la- bel is worth a thousand images in dataset distillation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2

    2024

  22. [30]

    Active learning for convolu- tional neural networks: A core-set approach

    Ozan Sener and Silvio Savarese. Active learning for convolu- tional neural networks: A core-set approach. InInternational Conference on Learning Representations, 2018. 2

    2018

  23. [31]

    Elucidating the design space of dataset condensation

    Shitong Shao, Zikai Zhou, Huanran Chen, and Zhiqiang Shen. Elucidating the design space of dataset condensation. arXiv preprint arXiv:2404.13733, 2024. 1, 2, 3, 4, 5, 6

    2024 arXiv

  24. [32]

    Generalized large-scale data condensation via var- ious backbone and statistical matching

    Xindong Zhang Shitong Shao, Zeyuan Yin and Zhiqiang Shen. Generalized large-scale data condensation via var- ious backbone and statistical matching. arXiv preprint arXiv:2311.17950, 2023. 2, 4, 5, 6

    2023 arXiv

  25. [33]

    UCF101: A dataset of 101 human actions classes from videos in the wild

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 2, 5

    2012 arXiv

  26. [34]

    Generative teaching networks: Accelerating neural architecture search by learning to gen- erate synthetic training data

    Felipe Petroski Such, Aditya Rawal, Joel Lehman, Kenneth Stanley, and Jeffrey Clune. Generative teaching networks: Accelerating neural architecture search by learning to gen- erate synthetic training data. In Proceedings of the Inter- national Conference on Machine Learning (I...

    2020

  27. [35]

    On the diversity and realism of distilled dataset: An efficient dataset distilla- tion paradigm

    Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distilla- tion paradigm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 6, 7

    2024

  28. [36]

    An empirical study of example forget- ting during deep neural network learning

    Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geof- frey J Gordon. An empirical study of example forget- ting during deep neural network learning. arXiv preprint arXiv:1812.05159, 2018. 2

    2018 arXiv

  29. [37]

    Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, pages 10078–10093. Curran Associates, Inc., 2022. 3

    2022

  30. [38]

    Learning spatiotemporal features with 3d convolutional networks

    Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torre- sani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. InICCV, pages 4489–4497,

  31. [39]

    A closer look at spatiotemporal convolutions for action recognition

    Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, pages 6450– 6459, 2018. 2

    2018

  32. [40]

    Cafe: Learning to condense dataset by aligning features

    Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In CVPR, 2022. 1, 2

    2022

  33. [41]

    Emphasizing dis- criminative features for dataset distillation in complex sce- narios

    Kai Wang, Zekai Li, Zhi-Qi Cheng, Samir Khaki, Ahmad Sajedi, Ramakrishna Vedantam, Konstantinos N Plataniotis, Alexander Hauptmann, and Yang You. Emphasizing dis- criminative features for dataset distillation in complex sce- narios. In Proceedings of the IEEE/CVF Conference on...

    2025

  34. [42]

    Action- clip: A new paradigm for video action recognition

    Mengmeng Wang, Jiazheng Xing, and Yong Liu. Action- clip: A new paradigm for video action recognition. CoRR, abs/2109.08472, 2021. 3

    2021 arXiv

  35. [43]

    Dataset dis- tillation with neural characteristic function: A minmax per- spective

    Shaobo Wang, Yicun Yang, Zhiyuan Liu, Chenghao Sun, Xuming Hu, Conghui He, and Linfeng Zhang. Dataset dis- tillation with neural characteristic function: A minmax per- spective. arXiv preprint arXiv:2502.20653, 2025. 8

    2025 arXiv

  36. [44]

    Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A. Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 5

    2018 arXiv

  37. [45]

    Dancing with still images: Video distillation via static-dynamic dis- entanglement, 2024

    Ziyu Wang, Yue Xu, Cewu Lu, and Yong-Lu Li. Dancing with still images: Video distillation via static-dynamic dis- entanglement, 2024. 1, 2, 3, 5, 6, 7, 8, 9

    2024

  38. [46]

    Vision-language dataset distillation

    Xindi Wu, Byron Zhang, Zhiwei Deng, and Olga Rus- sakovsky. Vision-language dataset distillation. Transactions on Machine Learning Research, 2024. 1, 2

    2024

  39. [47]

    Does graph distillation see like vision dataset counterpart? In Pro- ceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023

    Beining Yang, Kai Wang, Qingyun Sun, Cheng Ji, Xingcheng Fu, Hao Tang, Yang You, and Jianxin Li. Does graph distillation see like vision dataset counterpart? In Pro- ceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2

    2023

  40. [48]

    Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective

    Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 4, 5

    2023

  41. [49]

    Cutmix: Regu- larization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international con- ference on computer vision, pages 6023–6032, 2019. 6

    2019

  42. [50]

    Two trades is not baf- fled: Condensing graph via crafting rational gradient match- ing

    Tianle Zhang, Yuchen Zhang, Kun Wang, Kai Wang, Beining Yang, Kaipeng Zhang, Wenqi Shao, Ping Liu, Joey Tianyi Zhou, and Yang You. Two trades is not baf- fled: Condensing graph via crafting rational gradient match- ing. arXiv preprint arXiv:2402.04924, 2024. 1, 2

    2024 arXiv

  43. [51]

    Dataset condensation with differ- entiable siamese augmentation

    Bo Zhao and Hakan Bilen. Dataset condensation with differ- entiable siamese augmentation. In ICML, 2021. 1, 2

    2021

  44. [52]

    Dataset condensation with distri- bution matching

    Bo Zhao and Hakan Bilen. Dataset condensation with distri- bution matching. In WACV, 2023. 2

    2023

  45. [53]

    Dataset condensation with gradient matching

    Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929, 2020. 1, 2

    2006 arXiv

  46. [54]

    Towards stable and storage- efficient dataset distillation: Matching convexified trajectory

    Wenliang Zhong, Haoyu Tang, Qinghai Zheng, Mingzhu Xu, Yupeng Hu, and Liqiang Nie. Towards stable and storage- efficient dataset distillation: Matching convexified trajectory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2

    2025

  47. [55]

    Dataset distillation using neural feature regression

    Yongchao Zhou, Ehsan Nezhadarya, and Jimmy Ba. Dataset distillation using neural feature regression. arXiv preprint arXiv:2206.00719, 2022. 1, 2 12

    2022 arXiv

Pith tools

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