Pith. sign in

REVIEW 4 major objections 5 minor 20 references

Data Augmentation For Small Object using Fast AutoAugment

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 A cheap augmentation search lifts small-object detection by 20 percent on DOTA.

desk verdict A cheap, sensible search over copy-paste parameters for small-object detection, but the paper never shows the search—rather than the copy-paste itself—is what helps. read the letter →

arxiv 2506.08956 v1 pith:VH3UDU2C submitted 2025-06-10 cs.CV cs.LG

DaeEun Yoon , Semin Kim , SangWook Yoo , Jongha Lee This is my paper
classification cs.CVcs.LG
keywords objectdetectionsmalldataaugmentationFastAutoAugmentcopy-pastepolicysearchDOTAdatasetBayesianoptimization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a cheap policy search can find the right copy-paste augmentation for small-object detection, where standard augmentations fail. It uses Fast AutoAugment's proxy—the loss of a model trained without augmentation on augmented validation data—to score candidate policies over three copy-paste operations, a paste probability p, and a paste count m. On the DOTA-v2.0 aerial benchmark, the searched policies improve small-object mAP by 20% with Faster R-CNN and 17% with RetinaNet, and improve overall mAP as well. The practical point is that tuning just two numbers, p and m, with a fast proxy can replace expensive manual augmentation design.

What carries the argument

The central machinery is the Fast AutoAugment density-similarity proxy: an augmentation policy is good if a model trained on unaugmented data has low loss on validation data augmented with that policy. The search space is the copy-paste family from prior work—copy one, several, or all small objects and paste them at non-overlapping random locations—parameterized by paste probability p and paste count m (1 to 3). The proxy makes policy search cheap because it avoids retraining the detector for each candidate policy, and Bayesian optimization (TPE) explores the three-dimensional space of (operation, p, m).

What would settle it

Run the same final training with several hand-picked or random (operation, p, m) policies and compare final small-object mAP; if any achieve the same gain as the searched policy, the search itself adds nothing. More directly, compute the correlation between proxy loss and final small-object mAP across all evaluated policies; a near-zero correlation would falsify the paper's central claim.

Watch

Extended reading notes

Core claim

The paper's discovery is that Fast AutoAugment, originally designed for classification, transfers to small-object detection when its search space is restricted to copy-paste operations. The search splits the training set into K folds, trains a detector on each fold without augmentation, and then uses Bayesian optimization to minimize that detector's loss on held-out augmented data, varying the operation type, p, and m. The top low-loss policies are collected, one is applied per training iteration, and the final detector is trained on augmented data. The resulting policies have p and m inversely proportional, and they yield the reported gains on DOTA-v2.0 valid.

Load-bearing premise

The load-bearing premise is that the Fast AutoAugment proxy loss—the loss of a no-augmentation model on augmented validation data—predicts final small-object detection accuracy, so the policies that minimize it are the ones that improve mAP.

Editorial extensions

If this is right

  • Small-object mAP on DOTA-v2.0 valid rises from 0.402 to 0.485 with Faster R-CNN and from 0.122 to 0.143 with RetinaNet, with overall mAP improving as well.
  • The searched policies transfer across two very different detector families, suggesting the inverse p-m relationship reflects a property of the data rather than of a single architecture.
  • Policy search cost stays low: the Fast AutoAugment proxy uses already-trained fold models instead of repeatedly retraining a child detector for each candidate policy.
  • Because only operation, p, and m are searched, the method can be reproduced and extended with little engineering overhead beyond a standard detection pipeline.

Reading between the lines

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

  • If the proxy loss is well correlated with small-object mAP, the same search could be applied to other small-object-heavy benchmarks such as medical or satellite imagery, but the optimal policy would need to be re-searched per dataset.
  • The inverse relationship between p and m suggests the product p×m (the expected number of pasted objects per image) is the effective knob; a direct test would fix the product and vary p and m separately.
  • The reported gains could partly come from copy-paste itself rather than from the search; comparing searched policies against a few hand-picked p,m pairs would separate the contribution of the search.
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 manuscript proposes a data-augmentation search method for small-object detection in aerial images. It combines copy-paste augmentation from Kisantal et al. with the Fast AutoAugment search framework, using Bayesian optimization (TPE) to select the copy-paste operation, the paste probability p, and the paste count m. The method is evaluated on DOTA-v2.0 with Faster R-CNN and RetinaNet detectors, and the paper reports relative gains over a no-search baseline, including a 20% improvement in small-object mAP in Table 2.

Significance. If the central claim is confirmed, the method offers a low-cost way to specialize copy-paste augmentation for small aerial objects, building on Fast AutoAugment's efficiency rather than expensive policy search. The paper's strength is the simplicity of the pipeline: it reuses well-established components, namely copy-paste augmentation, Fast AutoAugment, and TPE, and targets a practically relevant regime where small-object performance lags. However, the current evidence is not yet sufficient: there are no equal-budget controls, no validation of the search proxy, no error bars, and no code or full training details, so the reported gains cannot yet be attributed to the policy search.

major comments (4)
  1. [§4, Tables 2 and 3] The reported gains are not attributable to the search because there is no control that trains the same final model for the same budget with (a) no augmentation, (b) a fixed copy-paste policy such as p=0.5 and m=1, or (c) random search over the same policy space. The baseline-to-SOA jump in Table 2, from 0.402 to 0.461 small-object mAP, already shows that the underlying copy-paste augmentation alone is effective; without an equal-budget control, the additional gain to 0.485 cannot be assigned to the Fast AutoAugment proxy optimization.
  2. [§3.2, Algorithm 1 lines 5-7] The load-bearing assumption is that the loss of a no-augmentation model on augmented validation data D_A predicts final small-object mAP. The paper never validates this proxy. I ask for a correlation or rank-agreement study: train several candidate policies, record both the proxy loss and the final small-object mAP, and show that the proxy orders the policies correctly. Without this, the policy selected by Bayesian optimization could simply be a low-loss augmentation with no downstream detection benefit.
  3. [§4, Tables 2 and 3] All results appear to be single runs on the DOTA-v2.0 valid split, with no error bars, no repeated seeds, and no test-dev evaluation. Given the stochasticity of both detector training and policy search, the abstract's 20% small-object improvement rests on a single validation-split measurement. Please report the mean and standard deviation over at least three seeds per configuration, and separate search variance from final training variance; if test-dev submission is possible, report those numbers or state that they are not available.
  4. [Algorithm 1 lines 6-8] The pseudo-code is under-specified to the point of not being reproducible. Line 6 does not state which validation images are used, whether the loss is detection loss or classification loss, or how the loss is aggregated over images. Line 7, T^k_t = T^k_t ∪ T_t, is circular before T^k_t is initialized, and line 8's 'select top N policies in T^k_t' mixes Tt with a set whose contents are never precisely defined. Please rewrite the algorithm with initialized variables, explicit loss definitions, and exact selection criteria, or provide the actual search configuration in the text.
minor comments (5)
  1. [Abstract and §2.1] The manuscript at times says 'Fast R-CNN' when the experiments use Faster R-CNN; please use consistent naming throughout.
  2. [Throughout] There are frequent typos and formatting issues, including 'RGBSshift', 'transform authorization', 'T able', and 'the prestigious Fast R-CNN'; a careful proofread is needed.
  3. [Fig. 4] The y-axis is described as 'the sum of the parameters p and m', but p is a probability and m is a count, so their sum is not a well-defined quantity; please plot p and m separately or define the exact aggregated quantity.
  4. [§3.3 and §4] The specific searched policies (operation, p, m) are not reported, which makes it hard to interpret Fig. 4 and to reproduce the method; consider listing the top policies in a table or appendix.
  5. [§4] The text reports a '9% increase' and a '20% improvement' without clarifying that these are relative changes in mAP and small-object mAP respectively; state absolute and relative values to avoid ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the search objective is a proxy loss on a training fold, while reported gains are measured on the held-out DOTA-v2.0 valid split.

full rationale

The paper's derivation chain is not circular. The claimed contribution is a Fast AutoAugment-style search over copy-paste operations, paste probability p, and paste count m, where policies are selected by minimizing the loss of an unaugmented model on augmented validation data (Algorithm 1, lines 5-7). The reported improvements in Tables 2 and 3 are AP values on the separate DOTA-v2.0 valid subset, not the search loss. No equation defines the reported mAP_S in terms of the search loss, and no fitted parameter is renamed as a prediction. The copy-paste search space is taken from prior work [13], but that is a normal use of an existing augmentation technique, not a self-citation or an imported uniqueness theorem. The main weakness is that the paper does not compare against an equal-budget random or fixed-policy baseline, so the 20% gain cannot be cleanly attributed to the Bayesian optimization over the proxy; however, this is an experimental control issue and an unvalidated proxy assumption, not circularity. The inverse-proportionality observation in Fig. 4 is a post-hoc summary without error bars, but it is not load-bearing for the central claim. Therefore the paper is self-contained against its stated evaluation and receives a circularity score of 0.

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

No invented entities or theoretical constructs. The method depends on two external assumptions about copy-paste efficacy and the Fast AutoAugment proxy, plus the transferability of fold-selected policies. Free parameters are p, m, and the undisclosed search budget.

free parameters (3)
  • copy-paste probability p = not tabulated; top-20 distribution shown in Fig. 4
    Optimized per operation by Bayesian optimization TPE in Sections 3.3-3.4; central to policy quality.
  • paste count m = searched over values 1 to 3
    Number of pastes per object; optimized jointly with p in Section 3.4.
  • search budget K, numSearch, N = not reported
    Kfold split size, number of TPE iterations, and top-N policies are set by the authors but undisclosed; they affect policy quality and search cost.
assumptions (3)
  • domain assumption Copy-pasting small objects into new locations improves small-object detection.
    Adopted from SOA [13] without re-derivation; if this premise fails, the policy search has no effect. Section 3.1.
  • domain assumption The Fast AutoAugment proxy loss, loss of a no-augmentation model on augmented validation data, is a valid selector for final detection mAP.
    Core search mechanism stated in Section 3.2 and Algorithm 1 line 6; not validated against final mAP in the paper.
  • domain assumption Policies selected on D_A, a fold of the training set, transfer to the final model trained on full D and evaluate on DOTA-v2.0 valid.
    Algorithm 1 lines 1-9; no analysis of distribution shift or fold sensitivity is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data Augmentation For Small Object using Fast AutoAugment." pith.science (2026). https://pith.science/paper/VH3UDU2C

@misc{pith2026250608956,
  author       = {Pith},
  title        = {Pith review of: Data Augmentation For Small Object using Fast AutoAugment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VH3UDU2C}},
  note         = {Machine review of arXiv:2506.08956}
}
read the original abstract

In recent years, there has been tremendous progress in object detection performance. However, despite these advances, the detection performance for small objects is significantly inferior to that of large objects. Detecting small objects is one of the most challenging and important problems in computer vision. To improve the detection performance for small objects, we propose an optimal data augmentation method using Fast AutoAugment. Through our proposed method, we can quickly find optimal augmentation policies that can overcome degradation when detecting small objects, and we achieve a 20% performance improvement on the DOTA dataset.

Figures

Figures reproduced from arXiv: 2506.08956 by the authors.

Figure 1
Figure 1. An inference sample of Faster R-CNN in MS COCO. The first row is an image consisting of a large object and a medium object instance, and the second row is an image consisting of a small object instance. Despite its clear visibility, small objects have lower detection performance compared to large object detection performance. that should be detected in the real world are often represented by small pixels in the imag… view at source ↗
Figure 2
Figure 2. Samples from DOTA. It consists of Google Earth, satellite, and aerial images. to another image patch. Similarly, it improves overall training performance but does not contribute to improving small object detection performance. The third is the absence of an optimal augmentation policy. Research on the augmenta￾tion method of small objects is being conducted steadily. However, most studies do not apply optimal augmen… view at source ↗
Figure 3
Figure 3. An overall procedure of augmentation policy search by Fast AutoAugment algorithm. 3.2 Searching Policies Searching for the optimal augmentation policy is based on Fast AutoAugment. Fast AutoAugment is a method of searching for an augmentation policy that is most suitable for the characteristics of Dataset by estimating density similarity between original data and augmented data. The methodology proposed by Fast Auto… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The distribution of probability and magnitude of the top 20 policies using Faster R-CNN, where the x-axis represents the type of copy-pasting and the y-axis represents the sum of the parameters p and m. Examining the parameter values, it can be observed that they exhib…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 11 canonical work pages

  1. [13]

    Augmentation for small object detection

    Mate Kisantal, Zbigniew Wojna, Jakub Murawski, Jacek Naruniec, and Kyunghyun Cho. Augmentation for small object detection. arXiv preprint arXiv:1902.07296, 2019

    arXiv 1902

  2. [16]

    Fast autoaugment

    Sungbin Lim, Ildoo Kim, Taesup Kim, Chiheon Kim, and Sungwoong Kim. Fast autoaugment. Advances in Neural Information Processing Systems, 32, 2019

    work page 2019

  3. [1]

    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. InEuropean conference on computer vision, pages 740–755. Springer, 2014

    2014

  4. [2]

    Random erasing data augmentation

    Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 13001–13008, 2020

    2020

  5. [3]

    Improved regularization of convolutional neural networks with cutout.arXiv preprint arXiv:1708.04552, 2017

    Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout.arXiv preprint arXiv:1708.04552, 2017

    arXiv 2017

  6. [4]

    mixup: Beyond empirical risk minimization.arXiv preprint arXiv:1710.09412, 2017

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

    arXiv 2017

  7. [5]

    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. InProceedings of the IEEE/CVF international conference on computer vision, pages 6023–6032, 2019

    2019

  8. [6]

    Faster r-cnn: Towards real-time object detection with region proposal networks.Advances in neural in- formation processing systems, 28, 2015

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks.Advances in neural in- formation processing systems, 28, 2015

    work page 2015

Show all 20 references
  1. [7]

    Object detection in aerial images: A large-scale benchmark and challenges.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, pages 1–1, 2021

    Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Object detection in aerial images: A large-scale benchmark and challenges.IEEE Trans- actions on Pattern Analysis and Machine Inte...

    2021

  2. [8]

    Rich feature hierarchiesforaccurateobjectdetectionandsemanticsegmentation

    Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchiesforaccurateobjectdetectionandsemanticsegmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014

    2014

  3. [9]

    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

  4. [10]

    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. In European conference on computer vision, pages 21–37. Springer, 2016

    2016

  5. [11]

    Scale-transferrable object detection

    Peng Zhou, Bingbing Ni, Cong Geng, Jianguo Hu, and Yi Xu. Scale-transferrable object detection. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 528–537, 2018

    2018

  6. [12]

    Stdnet: A convnet for small target detection

    Brais Bosquet, Manuel Mucientes, and Víctor M Brea. Stdnet: A convnet for small target detection. InBMVC, page 253, 2018

    2018

  7. [14]

    Autoaugment: Learning augmentation strategies from data

    Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113– 123, 2019

    2019

  8. [15]

    Population based augmentation: Efficient learning of augmentation policy schedules

    Daniel Ho, Eric Liang, Xi Chen, Ion Stoica, and Pieter Abbeel. Population based augmentation: Efficient learning of augmentation policy schedules. InInternational Conference on Machine Learning, pages 2731–2741. PMLR, 2019

    2019

  9. [17]

    Algorithms for hyper-parameter optimization.Advances in neural information processing systems, 24, 2011

    James Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. Algorithms for hyper-parameter optimization.Advances in neural information processing systems, 24, 2011

    2011

  10. [18]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python.Journal of Machine Learning ...

    2011

  11. [19]

    Ray: A distributed framework for emerging {AI} applications

    Philipp Moritz, Robert Nishihara, Stephanie Wang, Alexey Tumanov, Richard Liaw, Eric Liang, Melih Elibol, Zongheng Yang, William Paul, Michael I Jordan, et al. Ray: A distributed framework for emerging {AI} applications. In 13th USENIX Symposium on Operating Systems Design and...

    2018

  12. [20]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. InProceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017

    2017

Pith tools

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