Research map / Architectures and efficiency / CNN design
DropBlock: A regularization method for convolutional networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 structured DropBlock rather than no feature dropout | ResNet-50 · ImageNet validation | top-1 accuracy: 76.51 $\pm$ 0.07 -> 78.13 $\pm$ 0.05 | Helped |
| Add DropBlock instead of SpatialDropout | ResNet-50 · ImageNet validation | top-1 accuracy: 77.41 $\pm$ 0.04 -> 78.13 $\pm$ 0.05 | Helped |
| Add DropBlock and label smoothing together | ResNet-50 · ImageNet validation | top-1 accuracy: 78.13 $\pm$ 0.05 -> 78.35 $\pm$ 0.05 | Helped |
| Add DropBlock | AmoebaNet-B (6, 256) · ImageNet validation | top-1 accuracy: 82.25 -> 82.52 | Helped |
| Add DropBlock with block_size=5 | RetinaNet, ResNet-50 FPN · COCO val2017 | AP: 36.8 -> 38.4 | Helped |
| Add DropBlock (kp=0.2, block_size=16) | ResNet-101 FPN segmentation model · PASCAL VOC 2012 test | mIOU: 47.2 -> 53.4 | Helped |
| Use block_size=3 rather than block_size=5 (ablation) | RetinaNet + DropBlock · COCO val2017 | AP: 38.3 -> 38.4 | No clear effect |
| Use block_size=7 rather than block_size=5 (ablation) | RetinaNet + DropBlock · COCO val2017 | AP: 38.4 -> 38.2 | Hurt |
| Use block_size=16 rather than block_size=1 (ablation) | Segmentation model + DropBlock · PASCAL VOC 2012 test | mIOU: 51.0 -> 53.4 | Helped |
| Apply DropBlock in skip connections in addition to convolution layers (ablation) | ResNet-50 · ImageNet validation | top-1 accuracy | Helped |
| Gradually decrease keep_prob from 1 to the target value during training (ablation) | ResNet-50 + DropBlock · ImageNet validation | top-1 accuracy | Helped |
| Use a fixed keep_prob throughout training instead of a schedule (ablation) | ResNet-50 + DropBlock · ImageNet validation | top-1 accuracy | Hurt |
Closest papers by meaning
- AutoDropout: Learning Dropout Patterns to Regularize Deep Networks (AAAI 2021 · similarity 0.90)
- On the Regularization Properties of Structured Dropout (CVPR 2020 · similarity 0.89)
- Dropout with Expectation-linear Regularization (ICLR 2017 · similarity 0.89)
- Beyond Random Masking: When Dropout meets Graph Convolutional Networks (ICLR 2025 · similarity 0.89)
- BlockDrop: Dynamic Inference Paths in Residual Networks (CVPR 2018 · similarity 0.88)
- Jumpout : Improved Dropout for Deep Neural Networks with ReLUs (ICML 2019 · similarity 0.88)
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.