Pith. sign in

REVIEW 4 major objections 6 minor 65 references

Extending Dataset Pruning to Object Detection: A Variance-based Approach

T0 review · 4 major / 6 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 Dataset pruning works for object detection when images are scored by per-object variance.

desk verdict A useful variance-based detection pruning method that is overclaimed and missing its most relevant baseline, but the core idea is coherent and worth refereeing. read the letter →

arxiv 2505.17245 v1 pith:O74HBC6C submitted 2025-05-22 cs.CV cs.LG

Ryota Yagi This is my paper
classification cs.CVcs.LG
keywords datasetpruningcoresetselectionobjectdetectionVariance-basedPredictionScoretrainingdynamicsmeanAveragePrecisiondataefficiency
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 pruning has worked well for image classification but has rarely been carried over to object detection, and this paper claims to be the first principled extension of classification pruning techniques to that domain. The paper attempts to make the transfer by treating pruning as three linked problems: assigning each ground-truth object its best-matching prediction, scoring objects, and aggregating object scores into an image score. Its central proposal is the Variance-based Prediction Score (VPS), which ranks images by how much per-object Intersection-over-Union and confidence values fluctuate across training epochs. On PASCAL VOC and MS COCO, the paper reports that VPS-based pruning, particularly IoU variance with max aggregation, beats random selection, Forgetting, EL2N, AUM, and other baselines in mean Average Precision at pruning rates up to 90 percent. A fair reader would take the contribution to be a template for extending scoring-based data selection to structured prediction tasks.

What carries the argument

The load-bearing object is the per-object training-time signature: for each ground-truth box, CIPA records a time series of matched predictions, and VPS reduces that series to a single number, the standard deviation of IoU or confidence across epochs. The connection to informativeness is the observed moon-shaped relation in which low-variance objects are consistently easy or consistently hard, while mid-to-high variance objects sit at the boundary of learnability. Max aggregation then turns the object-level scores into an image-level ranking, and the main tables use max because the paper reports it as the best-performing choice.

What would settle it

Recompute VPS while excluding epochs in which an object had no matched prediction instead of carrying them into the variance: if the pruned-subset mAP advantage over Forgetting shrinks or reverses, the reported gain is an artifact of how missing predictions were handled.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the information useful for pruning a detection dataset lives at the level of individual object predictions rather than whole images. The Object-Level Attribution Problem is answered with Class-Prioritized IoU-Aware Prediction Assignment (CIPA), which for each ground-truth box picks the highest-IoU prediction of the same class at each training epoch. The Scoring Strategy Problem is answered with VPS, the standard deviation over training epochs of the matched prediction's IoU or confidence, and the Image-Level Aggregation Problem with a statistical aggregation such as max, sum, or mean. The empirical claim is that selecting images with high IoU variance (VPSiou) yields higher mAP than selecting by difficulty, loss, or forgetting across VOC and COCO, and that this holds at aggressive pruning ratios. The paper also argues that informative-sample selection matters more than annotation count or class-distribution balance.

Load-bearing premise

The ranking can only carry the argument if the per-object IoU and confidence traces are complete and if the max aggregation function was not chosen after inspecting the reported numbers; the paper leaves unspecified how epochs with no matched prediction enter the variance computation.

Editorial extensions

If this is right

  • If VPS is correct, detection datasets can be cut to 70 to 90 percent of their images with a smaller drop in mean Average Precision than random or difficulty-based selection, directly lowering storage and training cost.
  • The same three-step decomposition of attribution, scoring, and aggregation can be applied to other structured-output tasks such as instance segmentation, keypoint detection, or multi-label classification.
  • Selection scores computed once with a two-stage detector transfer, at least partly, to a one-stage detector such as YOLOv5, so a pruning decision does not have to be re-derived from scratch for every architecture.
  • Variance-based selection implies that moderately difficult objects, not the hardest or easiest ones, carry the training signal worth keeping, which challenges difficulty-only pruning intuitions.

Reading between the lines

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

  • If the moon-shaped relation between per-object variance and learnability is stable, VPS could be computed from a short warm-up run and used to prune before a full training pass, shrinking the overhead of collecting statistics.
  • The paper's finding that class-distribution shift does not hurt performance suggests that class balance can be ignored in favor of object-level uncertainty; a direct test would be to combine VPS with a class-balance constraint and see whether mAP rises further.
  • The advantage of max aggregation hints that a single hard-to-learn object can dominate an image's training value; this could be tested by ablating the highest-variance object from selected images and measuring the resulting mAP change.
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. The paper proposes a dataset pruning method for object detection. It identifies three design problems—object-level attribution, scoring strategy, and image-level aggregation—and addresses them with a class-prioritized IoU-aware prediction assignment (CIPA), a variance-based prediction score (VPS) computed over IoU and confidence time series, and a statistical aggregation function. Experiments on PASCAL VOC and MS COCO with Faster R-CNN-C4/ResNet-50, plus a YOLOv5m cross-architecture check, report that VPS-based pruning consistently improves mAP over several classification-derived baselines. The paper also analyzes the effect of annotation count and class distribution shift.

Significance. If the empirical claims hold, the paper would be a useful step toward extending dataset pruning beyond image classification: the three-way problem decomposition is clear, the CIPA matching procedure is a sensible way to assign predictions to ground-truth objects, and the VPS score is a natural adaptation of training-dynamics-based scoring to detection outputs. The evaluation covers two standard benchmarks and includes a cross-architecture test, and Appendix D provides additional aggregation results. However, the central empirical conclusion is currently supported only against a restricted set of baselines, with no seed variance reported, so the claimed consistency is not yet firmly established. The work is therefore of moderate significance and would benefit from a strengthened comparison and a more precise statement of the method's scope.

major comments (4)
  1. [§2.2, §4.1, Tables 1–3] The comparison set omits the detection-specific coreset method [30] (Coreset Selection for Object Detection), which the paper cites as related work and whose setup (Faster R-CNN-C4/ResNet-50) is adopted in §4.2. Without this baseline, the abstract's claim of "consistently outperforms prior dataset pruning methods" and the §2.2 claim of being "the first to naturally extend traditional dataset pruning techniques to object detection" are not supported. Please add [30] to the main comparisons and, if the novelty claim is retained, reconcile it with the existence of [30] and [49].
  2. [§4.3, Tables 1–2] All reported numbers come from single runs with no error bars or significance tests. On COCO, the differences between VPSiou and Forgetting in mAP are 0.32, 0.27, 0.40, and 0.56 points at 60%, 70%, 80%, and 90% pruning, respectively; these margins are small enough that run-to-run variance could change the ranking. Please report means and standard deviations over at least three seeds, or provide a paired significance test, for the main tables, and qualify the "consistently outperforms" wording accordingly.
  3. [§3.2, §3.4, Algorithm 1, Eq. (2)] Algorithm 1 assigns p_ij = ∅ when no candidate prediction overlaps a ground-truth box, but Eq. (2) does not specify how such missing epochs enter the variance computation: are they skipped, zero-imputed, or excluded from the average? If objects are frequently undetected in early epochs, the treatment of these missing values can bias the VPS estimate and distort the image ranking. Please state the exact rule used and report the frequency of undefined matches in the score-collection phase.
  4. [§4.4, Appendix D, Table 6] The statement in §4.4 that "the superiority of our proposed methods remains consistent regardless of the aggregation method used" is contradicted by Table 6: under average aggregation at 50% pruning, VPSconf achieves 45.13 mAP, below AUM (45.60), EL2N (45.75), Forgetting (46.24), IoU (45.53), and Confidence (45.24). Since all main tables use max aggregation and the choice of max is not justified a priori, the reported "consistent" advantage of VPSconf is at least partly an artifact of post-hoc aggregation selection. Please present aggregation results transparently and either justify max before seeing the test results or validate the aggregation choice on a held-out split.
minor comments (6)
  1. [Table 1] The header mixes "map@75" and "map@50" instead of the consistent "mAP@75" and "mAP@50" used elsewhere.
  2. [References] References [20] and [21] appear to be the same paper (Har-Peled and Mazumdar, "On coresets for k-means and k-median clustering"); one should be removed or the two distinct entries clarified.
  3. [Figure 5] The correlation coefficients r are computed from 11 method-level points with no uncertainty estimates; please describe these correlations as exploratory and add confidence intervals or a note on their limited statistical power.
  4. [§4.2] The sentence "we use an equal number of statistics per object instance when calculating scores" is unclear; please specify whether this refers to a fixed number of epochs, a fixed number of sampled objects per image, or something else.
  5. [Table 1 caption and §4.1] The baseline names "IoU" and "Confidence" overlap with the VPS variants "VPSiou" and "VPSconf"; please rename the baselines (e.g., "Mean-IoU" and "Mean-Conf") to avoid confusion.
  6. [Appendix A.5] The note on excluding training-dynamics-based methods [10,24] is informative, but the reasons for exclusion (computational cost and limited epochs) should be surfaced in the main text, since without it the baseline list in §4.1 appears to omit relevant recent methods without explanation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: VPS is defined directly from per-epoch IoU/confidence statistics, and the claimed gains are benchmarked against external baselines rather than derived from the method's own outputs.

full rationale

The VPS score (Eq. 2) is explicitly a standard deviation of per-epoch IoU or confidence values for each matched object, with no parameters fitted to mAP and no term that presupposes the final detection performance. The selection pipeline (CIPA matching, object-level scoring, image-level aggregation) is a data-processing construction; the empirical claim that VPS selects better subsets is tested against Random, IDP, Loss, AUM, Entropy, EL2N, Forgetting, IoU, and Confidence baselines on VOC and COCO, which are external to the method's construction. The paper does cite prior variance-based pruning ideas [10, 24] and detection-specific pruning [30], but the citations are contextual rather than load-bearing: the VPS definition does not invoke any result from those papers as a premise, and no uniqueness theorem or fitted parameter is imported. The use of Faster R-CNN both to compute scores and in main evaluation is a standard dataset-pruning protocol; it is a possible limitation but not circular because the score is not defined in terms of the evaluation metric or the selected subset. The main empirical weakness, namely the omission of [30] from the comparison tables despite citing it as a geometry-based detection pruning method, undermines the unqualified 'outperforms prior dataset pruning methods' conclusion, but this is an incompleteness in the benchmark rather than a circular derivation. The same applies to the post-hoc choice of max aggregation and the unspecified handling of unmatched objects in CIPA; these are transparency issues, not self-referential reductions. No equation in the paper equates a prediction to an input by construction, and no load-bearing step relies on a self-citation chain.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities. Its free parameters are the epoch count and aggregation choice, both selected by the author. The core assumptions are domain-level beliefs about the informativeness of prediction variance and about the matching and aggregation procedures.

free parameters (2)
  • Number of epochs T for score collection = 17 for VOC, 12 for COCO
    Chosen by the authors based on training budget and dataset size; not derived from first principles.
  • Aggregation function A (max/sum/mean) = max used in main tables
    The choice of max aggregation appears to be made after observing results across aggregation functions (Figure 4, Table 6), making it a post-hoc selection.
assumptions (4)
  • domain assumption IoU and confidence variance across training epochs is a meaningful informativeness signal for object detection.
    The entire VPS method relies on this premise, asserted in Sections 3.4 and 3.5 but not proven from a theory of detection learning.
  • domain assumption CIPA's class-prioritized IoU matching correctly attributes predictions to ground-truth objects.
    Algorithm 1 defines the matching rule, but the paper does not analyze failure modes such as duplicate assignments or objects without any positive IoU candidate.
  • domain assumption Aggregating object-level scores with a simple statistical function preserves the information needed to rank images.
    Equation 3 assumes that max, sum, or mean over objects yields an effective image-level score; no theoretical justification is offered beyond empirical comparisons.
  • domain assumption Linearly scaling training iterations with pruning ratio gives a fair comparison across methods.
    Appendix A.4 describes this schedule, but it is not shown that all methods benefit equally from the adjusted iteration counts, which could confound results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Extending Dataset Pruning to Object Detection: A Variance-based Approach." pith.science (2026). https://pith.science/paper/O74HBC6C

@misc{pith2026250517245,
  author       = {Pith},
  title        = {Pith review of: Extending Dataset Pruning to Object Detection: A Variance-based Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O74HBC6C}},
  note         = {Machine review of arXiv:2505.17245}
}
read the original abstract

Dataset pruning -- selecting a small yet informative subset of training data -- has emerged as a promising strategy for efficient machine learning, offering significant reductions in computational cost and storage compared to alternatives like dataset distillation. While pruning methods have shown strong performance in image classification, their extension to more complex computer vision tasks, particularly object detection, remains relatively underexplored. In this paper, we present the first principled extension of classification pruning techniques to the object detection domain, to the best of our knowledge. We identify and address three key challenges that hinder this transition: the Object-Level Attribution Problem, the Scoring Strategy Problem, and the Image-Level Aggregation Problem. To overcome these, we propose tailored solutions, including a novel scoring method called Variance-based Prediction Score (VPS). VPS leverages both Intersection over Union (IoU) and confidence scores to effectively identify informative training samples specific to detection tasks. Extensive experiments on PASCAL VOC and MS COCO demonstrate that our approach consistently outperforms prior dataset pruning methods in terms of mean Average Precision (mAP). We also show that annotation count and class distribution shift can influence detection performance, but selecting informative examples is a more critical factor than dataset size or balance. Our work bridges dataset pruning and object detection, paving the way for dataset pruning in complex vision tasks.

Figures

Figures reproduced from arXiv: 2505.17245 by the authors.

Figure 1
Figure 1. Overview of the score assignment process for object detection. For each ground-truth [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization of VPS scores (IoU, confidence) for object in PASCAL VOC dataset. The [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Pruned samples with ground truth bounding box based on low VPS [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: mAP comparison of different aggrega￾tion methods (max, average, sum) on the PASCAL VOC [15] dataset across varying pruning rates [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Plots showing averaged mAP vs. number of annotations and JS Divergence under high [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Overall pipeline of Faster R-CNN [40]. The input image is first processed by a backbone network to extract feature maps. These features are then fed into a Region Proposal Network (RPN) and Region of Interest (RoI) Pooling to generate RoI features. The resulting featur…
Figure 7
Figure 7. Figure 7: Heatmaps of sample IoU at different pruning rates on PASCAL VOC [ [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Comparison of sample selections from PASCAL VOC [ [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Comparison of sample selections from MS COCO [ [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 39 canonical work pages

  1. [30]

    Coreset selection for object detection

    Hyun Lee, Seung Kim, Jung Lee, Jae Yoo, and Nojun Kwak. Coreset selection for object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7682–7691, 2024

    work page 2024

  2. [49]

    $a^2$-DP: Annotation-aware data pruning for object detection, 2025

    Zengran Wang, Yanan Zhang, Yunlong Qi, Jiaxin Chen, Zehua Fu, and Di Huang. $a^2$-DP: Annotation-aware data pruning for object detection, 2025. URL https://openreview.net/ forum?id=Pc94ncbkoo. 12

    work page 2025

  3. [1]

    Balancing feature similarity and label vari- ability for optimal size-aware one-shot subset selection

    Abhinab Acharya, Dayou Yu, Qi Yu, and Xumin Liu. Balancing feature similarity and label vari- ability for optimal size-aware one-shot subset selection. InForty-first International Conference on Machine Learning, 2024

    work page 2024

  4. [2]

    Agarwal, Sariel Har-Peled, and Kasturi R

    Pankaj K. Agarwal, Sariel Har-Peled, and Kasturi R. Varadarajan. Approximating extent measures of points.Journal of the ACM, 51(4):606–635, 2004

    work page 2004

  5. [3]

    Green Recommender Systems: Optimizing Dataset Size for Energy-Efficient Algorithm Performance

    Ardalan Arabzadeh, Tobias Vente, and Joeran Beel. Green recommender systems: Optimizing dataset size for energy-efficient algorithm performance. InProceedings of the 18th ACM Conference on Recommender Systems (RecSys 2024), 2024. URL https://arxiv.org/ abs/2410.09359. Presented at the International Workshop on Recommender Systems for Sustainability and So...

    work page Pith review arXiv 2024

  6. [4]

    Smaller core-sets for balls

    Mihai Badoiu and Kenneth L Clarkson. Smaller core-sets for balls. InSODA, volume 3, pages 801–802, 2003

    work page 2003

  7. [5]

    Anchor pruning for object detection.Computer Vision and Image Understanding, 221:103445, 2022

    Matthias Bonnaerens, Matthias Freiberger, and Joni Dambre. Anchor pruning for object detection.Computer Vision and Image Understanding, 221:103445, 2022

    work page 2022

  8. [6]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean Conference on Computer Vision (ECCV), pages 213–229, Cham, 2020. Springer International Publishing

    work page 2020

Show all 65 references
  1. [7]

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

    2022

  2. [8]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3739–3748, 2023

    2023

  3. [9]

    Curriculum coarse-to- fine selection for high-ipc dataset distillation.arXiv preprint arXiv:2503.18872, 2025

    Yanda Chen, Gongwei Chen, Miao Zhang, Weili Guan, and Liqiang Nie. Curriculum coarse-to- fine selection for high-ipc dataset distillation.arXiv preprint arXiv:2503.18872, 2025

    2025 arXiv

  4. [10]

    Lightweight dataset pruning without full training via example difficulty and prediction uncertainty, 2025

    Yeseul Cho, Baekrok Shin, Changmin Kang, and Chulhee Yun. Lightweight dataset pruning without full training via example difficulty and prediction uncertainty, 2025. URL https: //arxiv.org/abs/2502.06905

    2025 arXiv

  5. [11]

    Selection via proxy: Efficient data selection for deep learning.arXiv preprint arXiv:1906.11829, 2019

    Cody Coleman, Christopher Yeh, Stephen Mussmann, Baharan Mirzasoleiman, Peter Bailis, Percy Liang, Jure Leskovec, and Matei Zaharia. Selection via proxy: Efficient data selection for deep learning.arXiv preprint arXiv:1906.11829, 2019

    1906 arXiv

  6. [12]

    Training-free dataset pruning for instance segmentation

    Yalun Dai, Lingao Xiao, Ivor Tsang, and Yang He. Training-free dataset pruning for instance segmentation. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=rvxWEbTtRY

    2025

  7. [13]

    Minimizing the accumulated trajectory error to improve dataset distillation

    Jiawei Du, Yidi Jiang, Vincent YF Tan, Joey Tianyi Zhou, and Haizhou Li. Minimizing the accumulated trajectory error to improve dataset distillation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749–3758, 2023

    2023

  8. [14]

    Glam: Efficient scaling of language models with mixture-of-experts

    Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, et al. Glam: Efficient scaling of language models with mixture-of-experts. InInternational conference on machine learning, pages 5547–5569. PMLR, 2022

    2022

  9. [15]

    The pascal visual object classes (voc) challenge.International journal of computer vision, 88: 303–338, 2010

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge.International journal of computer vision, 88: 303–338, 2010

    2010

  10. [16]

    Springer Science & Business Media, 2009

    Reza Zanjirani Farahani and Masoud Hekmatfar.Facility Location: Concepts, Models, Algo- rithms and Case Studies. Springer Science & Business Media, 2009. 10

    2009

  11. [17]

    Fast r-cnn

    Ross Girshick. Fast r-cnn. InProceedings of the IEEE international conference on computer vision, pages 1440–1448, 2015

    2015

  12. [18]

    Rich feature hierarchies for accurate object detection and semantic segmentation

    Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014

    2014

  13. [19]

    Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding

    Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. InInternational Conference on Learning Representations (ICLR), 2016

    2016

  14. [20]

    On coresets for k-means and k-median clustering

    Sariel Har-Peled and Soham Mazumdar. On coresets for k-means and k-median clustering. In Proceedings of the 36th Annual ACM Symposium on Theory of Computing, pages 291–300. ACM, 2004

    2004

  15. [21]

    On coresets for k-means and k-median clustering

    Sariel Har-Peled and Soham Mazumdar. On coresets for k-means and k-median clustering. In Proceedings of the thirty-sixth annual ACM symposium on Theory of computing, pages 291–300, 2004

    2004

  16. [22]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 770–778, 2016

    2016

  17. [23]

    Girshick

    Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross B. Girshick. Mask r-cnn. InProceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2961–2969, 2017

    2017

  18. [24]

    Large-scale dataset pruning with dynamic uncertainty

    Muyang He, Shuo Yang, Tiejun Huang, and Bo Zhao. Large-scale dataset pruning with dynamic uncertainty. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7713–7722, 2024

    2024

  19. [25]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications.arXiv preprint arXiv:1704.04861, 2017

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications.arXiv preprint arXiv:1704.04861, 2017

    2017 arXiv

  20. [26]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference

    Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. InProceedings of the IEEE Conference on Computer Vision and...

    2018

  21. [27]

    Llm-based knowledge pruning for time series data analytics on edge-computing devices.arXiv preprint arXiv:2406.08765, 2024

    Ruibing Jin, Qing Xu, Min Wu, Yuecong Xu, Dan Li, Xiaoli Li, and Zhenghua Chen. Llm-based knowledge pruning for time series data analytics on edge-computing devices.arXiv preprint arXiv:2406.08765, 2024

    2024 arXiv

  22. [28]

    Yolov5.https://github.com/ultralytics/yolov5, 2020

    Glenn Jocher et al. Yolov5.https://github.com/ultralytics/yolov5, 2020

    2020

  23. [29]

    Parp: Prune, adjust and re-prune for self-supervised speech recognition.Advances in Neural Information Processing Systems, 34:21256–21272, 2021

    Cheng-I Jeff Lai, Yang Zhang, Alexander H Liu, Shiyu Chang, Yi-Lun Liao, Yung-Sung Chuang, Kaizhi Qian, Sameer Khurana, David Cox, and Jim Glass. Parp: Prune, adjust and re-prune for self-supervised speech recognition.Advances in Neural Information Processing Systems, 34:21256...

    2021

  24. [31]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InComputer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings,...

    2014

  25. [32]

    Self-supervised learning for object detection: A survey

    Shuyang Liu, Xialei Liu, Yi Yang, and Cewu Lu. Self-supervised learning for object detection: A survey. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7770–7780, 2021. 11

    2021

  26. [33]

    Ssd: Single shot multibox detector

    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. InComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14, pa...

    2016

  27. [34]

    D2 pruning: Message passing for balancing diversity and difficulty in data pruning.arXiv preprint arXiv:2310.07931, 2023

    Adyasha Maharana, Prateek Yadav, and Mohit Bansal. D2 pruning: Message passing for balancing diversity and difficulty in data pruning.arXiv preprint arXiv:2310.07931, 2023

    2023 arXiv

  28. [35]

    Deep learning on a data diet: Finding important examples early in training.Advances in neural information processing systems, 34:20596–20607, 2021

    Mansheej Paul, Surya Ganguli, and Gintare Karolina Dziugaite. Deep learning on a data diet: Finding important examples early in training.Advances in neural information processing systems, 34:20596–20607, 2021

    2021

  29. [36]

    Identifying mislabeled data using the area under the margin ranking.Advances in Neural Information Processing Systems, 33:17044–17056, 2020

    Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying mislabeled data using the area under the margin ranking.Advances in Neural Information Processing Systems, 33:17044–17056, 2020

    2020

  30. [37]

    Fetch and forge: Efficient dataset condensation for object detection.Advances in Neural Information Processing Systems, 37:119283–119300, 2024

    Ding Qi, Jian Li, Jinlong Peng, Bo Zhao, Shuguang Dou, Jialin Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Fetch and forge: Efficient dataset condensation for object detection.Advances in Neural Information Processing Systems, 37:119283–119300, 2024

    2024

  31. [38]

    Yolov3: An incremental improvement, 2018

    Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018. URL https: //arxiv.org/abs/1804.02767

    2018 arXiv

  32. [39]

    You only look once: Unified, real-time object detection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 779–788, 2016

    2016

  33. [40]

    Faster R-CNN: Towards real-time object detection with region proposal networks

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. InAdvances in Neural Information Processing Systems, volume 28, pages 91–99, 2015

    2015

  34. [41]

    Svp-cf: Selection via proxy for collaborative filtering data.arXiv preprint arXiv:2107.04984, 2021

    Noveen Sachdeva, Carole-Jean Wu, and Julian McAuley. Svp-cf: Selection via proxy for collaborative filtering data.arXiv preprint arXiv:2107.04984, 2021

    2021 arXiv

  35. [42]

    Prototype selection for composite nearest neighbor classifiers

    David B Skalak. Prototype selection for composite nearest neighbor classifiers. 1997

    1997

  36. [43]

    Beyond neural scaling laws: beating power law scaling via data pruning.Advances in Neural Information Processing Systems, 35:19523–19536, 2022

    Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari Morcos. Beyond neural scaling laws: beating power law scaling via data pruning.Advances in Neural Information Processing Systems, 35:19523–19536, 2022

    2022

  37. [44]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5809–5818, 2024

    2024

  38. [45]

    Dataset cartography: Mapping and diagnosing datasets with training dynamics

    Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. Dataset cartography: Mapping and diagnosing datasets with training dynamics. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Proces...

    2020

  39. [46]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc V Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InInternational conference on machine learning, pages 6105–6114. PMLR, 2019

    2019

  40. [47]

    Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J. Gordon. An empirical study of example forgetting during deep neural network learning. InInternational Conference on Learning Representations (ICLR), 2019. URL https://open...

    2019

  41. [48]

    Dataset distillation

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

    2018 arXiv

  42. [50]

    Herding dynamical weights to learn

    Max Welling. Herding dynamical weights to learn. InProceedings of the 26th Annual Interna- tional Conference on Machine Learning, pages 1121–1128, 2009

    2009

  43. [51]

    CCNet: Extracting high quality monolingual datasets from web crawl data

    Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzmán, Armand Joulin, and Édouard Grave. CCNet: Extracting high quality monolingual datasets from web crawl data. InProceedings of the Twelfth Language Resources and Evalua- tion Conference, pa...

    2020

  44. [52]

    Detectron2

    Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github.com/facebookresearch/detectron2, 2019. Accessed: 2025-05-15

    2019

  45. [53]

    Moderate coreset: A universal method of data selection for real-world data-efficient deep learning

    Xiaobo Xia, Jiale Liu, Jun Yu, Xu Shen, Bo Han, and Tongliang Liu. Moderate coreset: A universal method of data selection for real-world data-efficient deep learning. InThe Eleventh International Conference on Learning Representations, 2022

    2022

  46. [54]

    Are large-scale soft labels necessary for large-scale dataset distil- lation? InThe Thirty-eighth Annual Conference on Neural Information Processing Systems,

    Lingao Xiao and Yang He. Are large-scale soft labels necessary for large-scale dataset distil- lation? InThe Thirty-eighth Annual Conference on Neural Information Processing Systems,

  47. [55]

    Dynamic data pruning for automatic speech recognition

    Qiao Xiao, Pingchuan Ma, Adriana Fernandez-Lopez, Boqian Wu, Lu Yin, Stavros Petridis, Mykola Pechenizkiy, Maja Pantic, Decebal Constantin Mocanu, and Shiwei Liu. Dynamic data pruning for automatic speech recognition. InProceedings of Interspeech 2024, Kos, Greece,

    2024

  48. [56]

    Data pruning can do more: A comprehensive data pruning approach for object re-identification.Transactions on Machine Learning Research, 2024

    Zi Yang, Haojin Yang, Soumajit Majumder, Jorge Cardoso, and Guillermo Gallego. Data pruning can do more: A comprehensive data pruning approach for object re-identification.Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openreview.net/ forum?id=vxxi7xzzn7

    2024

  49. [57]

    URLhttps://arxiv.org/abs/2406.18373

    ISCA. URLhttps://arxiv.org/abs/2406.18373

    arXiv

  50. [58]

    Amin Zareian, Yuval Atzmon, Eli Shechtman, and David A. Ross. Open-vocabulary object detection. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3335–3344, 2021

    2021

  51. [59]

    Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective.Advances in Neural Information Processing Systems, 36:73582–73603, 2023

    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, 36:73582–73603, 2023

    2023

  52. [60]

    Dataset condensation with differentiable siamese augmentation

    Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674–12685. PMLR, 2021

    2021

  53. [61]

    Spanning training progress: Temporal dual-depth scoring (tdds) for enhanced dataset pruning

    Xin Zhang, Jiawei Du, Yunsong Li, Weiying Xie, and Joey Tianyi Zhou. Spanning training progress: Temporal dual-depth scoring (tdds) for enhanced dataset pruning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26223–26232, 2024

    2024

  54. [62]

    Coverage-centric coreset selection for high pruning rates.arXiv preprint arXiv:2210.15809, 2022

    Haizhong Zheng, Rui Liu, Fan Lai, and Atul Prakash. Coverage-centric coreset selection for high pruning rates.arXiv preprint arXiv:2210.15809, 2022

    2022 arXiv

  55. [63]

    Boosting the cross-architecture gener- alization of dataset distillation through an empirical study.arXiv preprint arXiv:2312.05598, 2023

    Lirui Zhao, Yuxin Zhang, Fei Chao, and Rongrong Ji. Boosting the cross-architecture gener- alization of dataset distillation through an empirical study.arXiv preprint arXiv:2312.05598, 2023

    2023 arXiv

  56. [65]

    Deformable detr: Deformable transformers for end-to-end object detection

    Xizhou Zhu, Hengduo Li, Jiefeng Li, Zhi Tian, Bo Li, Xuanshi Li, and Wenyu Liu. Deformable detr: Deformable transformers for end-to-end object detection. InProceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1840–1849, 2021. 13 A Dataset and Imple...

    2021

  57. [2024]

    URLhttps://openreview.net/forum?id=12A1RT1L87

Pith tools

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