Research map / Detection and segmentation / Weakly supervised detection and segmentation
CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features
Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, Youngjoon Yoo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CutMix augmentation | ResNet-50 · ImageNet-1K val | top-1 error (%): 23.68 -> 21.40 | Improved |
| add CutMix augmentation | ResNet-101 · ImageNet-1K val | top-1 error (%): 21.87 -> 20.17 | Improved |
| add CutMix augmentation | PyramidNet-200 · CIFAR-100 | top-1 error (%): 16.45 -> 14.47 | Improved |
| add CutMix on top of ShakeDrop | PyramidNet-200 · CIFAR-100 | top-1 error (%): 15.08 -> 13.81 | Improved |
| apply CutMix at image level rather than feature level (ablation) | PyramidNet-200 · CIFAR-100 | top-1 error (%): 21.80 -> 21.40 | Improved |
| sample box centers uniformly instead of Gaussian (ablation) | CutMix · CIFAR-100 | top-1 error (%): 15.95 -> 14.47 | Improved |
| sample box size randomly instead of fixed 16x16 (ablation) | CutMix · CIFAR-100 | top-1 error (%): 14.97 -> 14.47 | Improved |
| mix labels by area ratio instead of one-hot of larger patch (ablation) | CutMix · CIFAR-100 | top-1 error (%): 15.89 -> 14.47 | Improved |
| mix labels by lambda instead of fixed 0.5/0.5 (ablation) | CutMix · CIFAR-100 | top-1 error (%): 15.17 -> 14.47 | Improved |
| apply CutMix with constant probability instead of linear schedule (ablation) | CutMix · CIFAR-100 | top-1 error (%): 14.72 -> 14.47 | Improved |
| pretrain backbone with Mixup | ResNet-50 Faster-RCNN · Pascal VOC07 detection | mAP: 75.6 -> 73.9 | Hurt |
| pretrain backbone with CutMix | ResNet-50 SSD · Pascal VOC07 detection | mAP: 76.7 -> 77.6 | Improved |
Later papers that compared with this method
12 later accepted papers report CutMix as a baseline in their results tables; 11 of them report a better number on most of the results they share with it.
- Learning from Synthetic Data via Provenance-Based Input Gradient Guidance (CVPR 2026 · better on 4 of 4 shared results)
- Multi-Perspective Data Augmentation for Few-shot Object Detection (ICLR 2025 · better on 1 of 1 shared result)
- Using Unreliable Pseudo-Labels for Label-Efficient Semantic Segmentation (IJCV 2024 · better on 1 of 1 shared result)
- Background-Aware Classification Activation Map for Weakly Supervised Object Localization (TPAMI 2023 · better on 3 of 3 shared results)
- Bagging Regional Classification Activation Maps for Weakly Supervised Object Localization (ECCV 2022 · better on 3 of 3 shared results)
- Semantic Novelty Detection via Relational Reasoning (ECCV 2022 · better on 2 of 2 shared results)
- Anti-Adversarially Manipulated Attributions for Weakly Supervised Semantic Segmentation and Object Localization (TPAMI 2022 · better on 2 of 2 shared results)
- CycleMix: A Holistic Strategy for Medical Image Segmentation From Scribble Supervision (CVPR 2022 · better on 1 of 1 shared result)
4 more not listed.
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Provable Benefit of Cutout and CutMix for Feature Learning (NeurIPS 2024 · similarity 0.89)
- SaliencyMix: A Saliency Guided Data Augmentation Strategy for Better Regularization (ICLR 2021 · similarity 0.87)
- DropBlock: A regularization method for convolutional networks (NeurIPS 2018 · similarity 0.86)
- Supervision Interpolation via LossMix: Generalizing Mixup for Object Detection and Beyond (AAAI 2024 · similarity 0.84)
- CutFreq: Cut-and-Swap Frequency Components for Low-Level Vision Augmentation (AAAI 2024 · similarity 0.84)
- Recognize Any Regions (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
