Pith. sign in

REVIEW 4 major objections 5 minor 35 references

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

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Video dataset distillation can preserve motion by gating gradient updates with inter-frame saliency, and the paper reports large gains over prior video distillation methods.

desk verdict A plausible and efficiently-motivated video dataset distillation framework with large reported gains, but the key temporal saliency mask is under-specified enough that the results cannot yet be reproduced from the paper alone. read the letter →

arxiv 2505.20694 v1 pith:3MU33RXD submitted 2025-05-27 cs.CV cs.LG

Xulin Gu , Xinhao Zhong , Zhixing Wei , Yimin Zhou , Shuoyang Sun , Bin Chen , Hongpeng Wang , Yuan Luo This is my paper
classification cs.CVcs.LG
keywords datasetdistillationvideotemporalsaliencyinter-framedifferencesyntheticdataactionrecognitionuni-levelframeworkaugmentation
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

Dataset distillation compresses a large video dataset into a tiny synthetic one that can still train action-recognition models. The paper tries to establish that this can be done for video without the excessive cost of prior two-stage methods, and that motion information does not have to be sacrificed. Its central proposal is a Temporal Saliency-Guided Filter that measures each frame's motion content from smoothed inter-frame pixel differences and uses that measurement to decide how strongly each frame should be optimized. If the claim holds, video dataset distillation becomes cheap enough for practical use while retaining the temporal cues that action recognition depends on.

What carries the argument

The central object is the Temporal Saliency-Guided Filter (TSGF), a frame-level mask computed as $M_i = \frac{\max(\epsilon - s_i, 0)}{\max(s) - \min(s)}$, where $s_i$ is the smoothed inter-frame difference of frame $i$. It has two uses: TSGF$_{\text{O}}$ multiplies the distillation gradient frame-wise before updating synthetic videos, so high-saliency frames receive smaller updates, and TSGF$_{\text{A}}$ uses the same saliency signal to apply VideoMix augmentation only to low-saliency frames. The filter is what turns a generic decoupled distillation objective into a video-aware one.

What would settle it

A direct test is to replace the inter-frame-difference saliency with optical-flow magnitude or a learned motion score and compare MiniUCF IPC=5 accuracy; if the reported 54.8% does not hold up when the saliency measure is changed while all else is fixed, then the specific heuristic is the source of the gain.

Watch

Extended reading notes

Core claim

The paper's central claim is that a uni-level optimization framework—train one video classifier on the real data, then optimize synthetic videos against that frozen model using cross-entropy and batch-normalization statistics—can outperform two-stage video distillation methods when augmented with a temporal saliency mask. The mask, built from smoothed inter-frame differences, scales down gradient updates for frames that move the most, and a matching augmentation stage applies VideoMix only to low-saliency frames. With this mechanism the paper reports 54.8% top-1 accuracy on MiniUCF at IPC=5 versus 33.3% for IDTD, 15.2% top-5 accuracy on SSv2 at IPC=5 versus 9.5% for IDTD, and consistent improvements across all tested settings with reduced distillation cost.

Load-bearing premise

The load-bearing premise is that smoothed inter-frame pixel differences are a valid measure of temporal saliency, and that high-saliency frames should receive smaller gradient updates to preserve motion.

Editorial extensions

If this is right

  • On MiniUCF at IPC=5, the method reaches 54.8% top-1 accuracy, 21.5 points above IDTD's 33.3%.
  • On SSv2 at IPC=5, it reaches 15.2% top-5 accuracy, up from 9.5% for IDTD, under a compression ratio below 1%.
  • Both TSGF components contribute: adding temporal saliency-guided optimization and augmentation raises MiniUCF IPC=5 accuracy from 40.5% to 54.8%.
  • The distilled datasets transfer to other architectures, including ConvNet3D, CNN+GRU, and CNN+LSTM, better than prior video distillation methods.
  • The paper reports substantially lower GPU memory and time than existing image- and video-distillation methods across IPC settings.

Reading between the lines

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

  • The same mask mechanism could be applied to the interpolation stage of two-stage video distillation methods, since TSGF is orthogonal to the distillation objective.
  • The saliency heuristic could be replaced by optical-flow magnitude or a learned motion score; the paper's own limitation section notes that optical flow is not used, and such alternatives may strengthen temporal compression.
  • The threshold $\epsilon$ and smoothing window in the saliency computation are likely dataset-sensitive, and a percentile-based or learned saliency threshold would make the method more portable.
  • Because real-frame initialization outperforms random noise by a large margin on MiniUCF at IPC=5, an experiment that disables TSGF while keeping real initialization would isolate how much of the gain comes from the filter itself rather than from the temporal priors in the initialization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a uni-level video dataset distillation framework that first trains a video classification teacher on the full dataset and then optimizes synthetic videos by matching classification loss and batch-normalization statistics. To preserve temporal information, it introduces a Temporal Saliency-Guided Filter (TSGF) with two components: TSGF-O, which uses smoothed inter-frame differences to compute a per-frame gradient mask, and TSGF-A, which applies VideoMix only to low-saliency frames during evaluation. Experiments on MiniUCF, HMDB51, Kinetics-400, and SSv2 report state-of-the-art accuracy compared with coreset, image-distillation, and video-distillation baselines, and an ablation on MiniUCF at IPC=5 shows that both TSGF components contribute (baseline 40.5%, +TSGF-A 46.9%, +TSGF-O 51.7%, both 54.8%). The paper also includes ablations on initialization, IPC, augmentation, static/dynamic groups, cross-architecture generalization, and visualizations of optical flow and frame differences.

Significance. If the empirical claims hold, the paper demonstrates a simple and comparatively cheap decoupled video distillation baseline that substantially outperforms prior two-stage methods, with the TSGF mechanism being orthogonal to existing distillation pipelines. The ablations in Table 3 and Appendix D provide useful evidence that temporal saliency information matters for video distillation. However, the central mechanism is currently under-specified: the threshold and smoothing parameters are not reported, no code is released, and the scalability evidence is qualitative rather than quantitative. These issues must be resolved before the state-of-the-art claim can be verified.

major comments (4)
  1. [§3.3 (Eqs. (6)–(10)) and Appendix F.1] The TSGF mechanism is not well-defined as written. Equation (7) defines saliency s_i with an unspecified window length k and weights α_k, and Eq. (8) and Eq. (10) both depend on a threshold ε that is never given. Appendix F.1 says that unlisted parameters "are assumed to follow the default values specified in our code," but no code is released and the paper does not report ε for any dataset or IPC setting. Because the mask M depends on the raw scale of s, ε cannot be interpreted without these values. The authors should supply the values of ε, k, and α, and a sensitivity analysis over ε (and ideally k) for the Table 3 ablation; without this, the reported gains from TSGF-O and TSGF-A cannot be reproduced or attributed to the mechanism rather than threshold tuning.
  2. [§3.3, Eq. (8)] The mask M = max(ε−s, 0)/(max(s)−min(s)) is undefined when max(s)=min(s), which occurs for videos with constant inter-frame difference, such as static clips in the static/dynamic split of Appendix D. The paper gives no fallback for this case. In addition, for any ε ≤ min(s) the mask is identically zero and the synthetic videos are frozen at initialization, while for ε ≥ max(s) the hard threshold degenerates into a soft reweighting. The manuscript should state how ε is chosen relative to the saliency distribution, what happens in the degenerate case, and whether the empirical results are robust to a range of ε.
  3. [§4.2, Figure 1, and Appendix F.2] The title and abstract promise a scalable framework, and Figure 1 is described as comparing "distillation cost," but neither the figure nor any table reports quantitative training time or GPU memory numbers. Appendix F.2 only lists the hardware. The reader cannot verify the claimed reductions in memory and time, which are part of the central contribution. I request a table with wall-clock distillation time and peak GPU memory for the compared methods across datasets and IPC settings.
  4. [§4.3, Table 2] The dagger in Table 2 states that the teacher model used by the authors on Kinetics-400 is less accurate than the teacher models of the baselines due to different hyper-parameter settings. Since the teacher is the sole source of supervision in the decoupled framework, this discrepancy needs explanation: what hyper-parameters differ, and how does the comparison look when all methods use the same teacher? The claim of consistent state-of-the-art performance across all settings depends on this comparison being fair.
minor comments (5)
  1. [§3.3, Eq. (6)] The first and last frames are undefined because f_{i−1} and f_{i+1} do not exist; please specify boundary handling.
  2. [Algorithm 1] The loop header "for k ← −0 to K − 1" should be "for k ← 0 to K − 1."
  3. [§2.1] In the discussion of image dataset distillation, DM is cited as [31], but the reference list identifies [31] as gradient matching (DC) and [30] as distribution matching; the citation should be corrected.
  4. [Throughout] Please correct typographical errors: "samppled" near Eq. (2), "extention" in Section 1, and "teporal" in Section 3.3.
  5. [Appendix F.1, Table 9] The hyperparameter table lists only lr and r_bn; even if ε and smoothing parameters are added to the main text, they should also appear in Table 9 for completeness.

Circularity Check

0 steps flagged · score 0.0 of 10

The distillation pipeline is a standard decoupled teacher-guided synthesis with held-out test evaluation; no load-bearing step reduces to its own inputs.

full rationale

The derivation chain is: train a teacher on the full training set (Eq. 3), synthesize a compact video set by matching the teacher's BN statistics and classification loss (Eqs. 4-5), modulate gradients with a hand-designed inter-frame-difference mask (Eqs. 6-9), then train fresh student models on the distilled set and report accuracy on the original test set (Section 4). The reported numbers are therefore measured on data never used in distillation, not on a quantity that appears in the optimization objective. The TSGF mask is a heuristic function of smoothed pixel differences; it contains no parameter fitted to test accuracy, and the paper does not claim to 'predict' inter-frame saliency from the distilled set. Eq. (8)'s unspecified epsilon and denominator are reproducibility concerns (the normalization max(s)-min(s) can degenerate and the threshold is not given), but they do not make the central SOTA claim an identity. The only self-citations ([33],[34],[35]) appear in the related-work enumeration of image-distillation methods and are not used to justify the TSGF mechanism or to rule out alternatives; no author-specific uniqueness theorem is invoked. The paper even states as a limitation that optical-flow-style motion compensation is not incorporated. Hence, the central derivation is self-contained against held-out benchmarks and no circular step is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method rests on standard decoupled-distillation assumptions plus an unvalidated temporal-saliency heuristic. It introduces no new physical or conceptual entities. Several central hyperparameters are unstated, and the temporal-saliency heuristic is the least externally supported piece of the framework.

free parameters (5)
  • epsilon (temporal saliency threshold) = not reported
    Used in Equations (8) and (10) to decide which frames receive constrained optimization and which frames are augmented. No value or tuning procedure is given anywhere in the paper.
  • smoothing window length k and weights alpha_k = not reported
    Equation (7) defines saliency as a weighted sum over previous frame differences using a window function, but the window type and length are never specified.
  • distillation iterations K = not reported
    Algorithm 1 loops for K iterations but K is never given, and it directly controls the amount of optimization.
  • learning rate lr = 0.25 for MiniUCF/HMDB51, 0.3 for Kinetics-400/SSv2
    Reported in Table 9; chosen per dataset and controls the synthetic video update step.
  • BN regularization coefficient r_bn = 0.001, 0.005, or 0.01 depending on dataset and IPC
    Reported in Table 9; weights the BN statistic matching term in Equation (5), which is central to the distillation objective.
assumptions (4)
  • domain assumption A pre-trained video classifier trained on the full dataset preserves sufficient class-discriminative and BN-statistical information to synthesize useful videos.
    Invoked in Section 3.2: the entire distillation relies on matching teacher cross-entropy and BN running statistics through Equations (3) to (5).
  • domain assumption Smoothed inter-frame pixel differences are a valid measure of temporal saliency or frame informativeness.
    Introduced in Section 3.3, Equations (6) and (7); no evidence or comparison with optical flow or learned motion features is provided.
  • ad hoc to paper Frames with higher temporal saliency should receive smaller gradient updates to preserve motion information.
    Equation (8) and (9) implement this inverse relationship, but the direction is asserted rather than derived or validated.
  • ad hoc to paper Augmenting only low-saliency frames with VideoMix preserves temporal coherence while increasing diversity.
    Equation (10) applies augmentation only when s_i <= epsilon; this design choice is not tested against other thresholds or augmentation policies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Temporal Saliency-Guided Distillation: A Scalable Framework for Distilling Video Datasets." pith.science (2026). https://pith.science/paper/3MU33RXD

@misc{pith2026250520694,
  author       = {Pith},
  title        = {Pith review of: Temporal Saliency-Guided Distillation: A Scalable Framework for Distilling Video Datasets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3MU33RXD}},
  note         = {Machine review of arXiv:2505.20694}
}
read the original abstract

Dataset distillation (DD) has emerged as a powerful paradigm for dataset compression, enabling the synthesis of compact surrogate datasets that approximate the training utility of large-scale ones. While significant progress has been achieved in distilling image datasets, extending DD to the video domain remains challenging due to the high dimensionality and temporal complexity inherent in video data. Existing video distillation (VD) methods often suffer from excessive computational costs and struggle to preserve temporal dynamics, as na\"ive extensions of image-based approaches typically lead to degraded performance. In this paper, we propose a novel uni-level video dataset distillation framework that directly optimizes synthetic videos with respect to a pre-trained model. To address temporal redundancy and enhance motion preservation, we introduce a temporal saliency-guided filtering mechanism that leverages inter-frame differences to guide the distillation process, encouraging the retention of informative temporal cues while suppressing frame-level redundancy. Extensive experiments on standard video benchmarks demonstrate that our method achieves state-of-the-art performance, bridging the gap between real and distilled video data and offering a scalable solution for video dataset compression.

Figures

Figures reproduced from arXiv: 2505.20694 by the authors.

Figure 1
Figure 1. Comparison of test accuracy and distillation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison between our TSGF framework and the traditional two-stage video distillation [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Experimental results on MiniUCF under IPC=1 with varying numbers of frames. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Experimental results on MiniUCF under IPC=1 with varying numbers of frames. Number of Frames [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Optical flows of MiniUCF. To qualitatively analyze how our method pre￾serves temporal dynamics, we visualize the op￾tical flow of distilled videos in [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Inter-frame difference visualization on MiniUCF. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 25 canonical work pages

  1. [1]

    Cisco visual networking index (vni) complete forecast update, 2017–2022

    Thomas Barnett, Shruti Jain, Usha Andra, and Taru Khurana. Cisco visual networking index (vni) complete forecast update, 2017–2022. Americas/EMEAR Cisco Knowledge Network (CKN) Presentation, 1(1), 2018

    work page 2017

  2. [2]

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

    Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299–6308, 2017

    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 Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 4750–4759, 2022

    work page 2022

  4. [4]

    Generalizing Dataset Distillation via Deep Generative Prior

    George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Generalizing dataset distillation via deep generative prior. arXiv preprint arXiv:2305.01649, 2023

    work page Pith review arXiv 2023

  5. [5]

    Influence- guided diffusion for dataset distillation

    Mingyang Chen, Jiawei Du, Bo Huang, Yi Wang, Xiaobo Zhang, and Wei Wang. Influence- guided diffusion for dataset distillation. In The Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  6. [6]

    A Large-Scale Study on Video Action Dataset Condensation

    Yang Chen, Sheng Guo, Bo Zheng, and Limin Wang. A large-scale study on video action dataset condensation. arXiv preprint arXiv:2412.21197, 2024

    work page Pith review arXiv 2024

  7. [7]

    Imagenet: A large- scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  8. [8]

    Diversity-Driven Synthesis: Enhancing Dataset Distillation through Directed Weight Adjustment

    Jiawei Du, Xin Zhang, Juncheng Hu, Wenxin Huang, and Joey Tianyi Zhou. Diversity-driven synthesis: Enhancing dataset distillation through directed weight adjustment. arXiv preprint arXiv:2409.17612, 2024

    work page Pith review arXiv 2024

Show all 35 references
  1. [9]

    something something

    Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In Pr...

    2017

  2. [10]

    Efficient dataset distillation via minimax diffusion

    Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15793–15803, 2024

    2024

  3. [11]

    Hmdb: a large video database for human motion recognition

    Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556–2563. IEEE, 2011

    2011

  4. [12]

    Dataset distillation with infinitely wide convolutional networks

    Timothy Nguyen, Roman Novak, Lechao Xiao, and Jaehoon Lee. Dataset distillation with infinitely wide convolutional networks. Advances in Neural Information Processing Systems , 34:5186–5198, 2021

    2021

  5. [13]

    Datadam: Efficient dataset distillation with attention matching

    Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstanti- nos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. InProceedings of the IEEE/CVF International Conference on Computer Vision , pages 17097–17107, 2023

    2023

  6. [14]

    Generalized large-scale data condensation via various backbone and statistical matching

    Shitong Shao, Zeyuan Yin, Muxin Zhou, Xindong Zhang, and Zhiqiang Shen. Generalized large-scale data condensation via various backbone and statistical matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 16709–16718, 2024

    2024

  7. [15]

    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. 10

    2024 arXiv

  8. [16]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. ICLR, 2021

    2021

  9. [17]

    A dataset of 101 human action classes from videos in the wild

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11):1–7, 2012

    2012

  10. [18]

    Dˆ 4: Dataset distillation via disentangled diffusion model

    Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. Dˆ 4: Dataset distillation via disentangled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5809–5818, 2024

    2024

  11. [19]

    On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm

    Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9390–9399, 2024

    2024

  12. [20]

    Learning spa- tiotemporal features with 3d convolutional networks

    Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spa- tiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489–4497, 2015

    2015

  13. [21]

    Dataset distillation

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

    2018 arXiv

  14. [22]

    Adaptive focus for efficient video recognition

    Yulin Wang, Zhaoxi Chen, Haojun Jiang, Shiji Song, Yizeng Han, and Gao Huang. Adaptive focus for efficient video recognition. In proceedings of the IEEE/CVF international conference on computer vision, pages 16249–16258, 2021

    2021

  15. [23]

    Dancing with still images: video distillation via static-dynamic disentanglement

    Ziyu Wang, Yue Xu, Cewu Lu, and Yong-Lu Li. Dancing with still images: video distillation via static-dynamic disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6296–6304, 2024

    2024

  16. [24]

    Dreaming to distill: Data-free knowledge transfer via deepinversion

    Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deepinversion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 871...

    2020

  17. [25]

    Dataset distillation via curriculum data synthesis in large data era

    Zeyuan Yin and Zhiqiang Shen. Dataset distillation via curriculum data synthesis in large data era. Transactions on Machine Learning Research, 2024

    2024

  18. [26]

    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. Advances in Neural Information Processing Systems , 36, 2024

    2024

  19. [27]

    Cutmix: Regularization strategy to train strong classifiers with localizable features

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

    2019

  20. [28]

    Videomix: Rethinking data augmentation for video classification

    Sangdoo Yun, Seong Joon Oh, Byeongho Heo, Dongyoon Han, and Jinhyung Kim. Videomix: Rethinking data augmentation for video classification. arXiv preprint arXiv:2012.03457, 2020

    2012 arXiv

  21. [29]

    mixup: Beyond empirical risk minimization

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017

    2017 arXiv

  22. [30]

    Dataset condensation with distribution matching

    Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. arXiv preprint arXiv:2110.04181, 2021

    2021 arXiv

  23. [31]

    Dataset condensation with gradient matching

    Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. ICLR, 1(2):3, 2021

    2021

  24. [32]

    Video set distilla- tion: Information diversification and temporal densification

    Yinjie Zhao, Heng Zhao, Bihan Wen, Yew-Soon Ong, and Joey Tianyi Zhou. Video set distilla- tion: Information diversification and temporal densification. arXiv preprint arXiv:2412.00111, 2024

    2024 arXiv

  25. [33]

    Going beyond feature similarity: Effective dataset distillation based on class-aware conditional mutual information

    Xinhao Zhong, Bin Chen, Hao Fang, Xulin Gu, Shu-Tao Xia, and En-Hui Yang. Going beyond feature similarity: Effective dataset distillation based on class-aware conditional mutual information. arXiv preprint arXiv:2412.09945, 2024. 11

    2024 arXiv

  26. [34]

    Hierarchical features matter: A deep exploration of gan priors for improved dataset distillation

    Xinhao Zhong, Hao Fang, Bin Chen, Xulin Gu, Tao Dai, Meikang Qiu, and Shu-Tao Xia. Hierarchical features matter: A deep exploration of gan priors for improved dataset distillation. arXiv preprint arXiv:2406.05704, 2024

    2024 arXiv

  27. [35]

    Efficient dataset distillation via diffusion-driven patch selection for improved generalization

    Xinhao Zhong, Shuoyang Sun, Xulin Gu, Zhaoyang Xu, Yaowei Wang, Jianlong Wu, and Bin Chen. Efficient dataset distillation via diffusion-driven patch selection for improved generalization. arXiv preprint arXiv:2412.09959, 2024. 12 Technical Appendices and Supplementary Material...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.