[BTY#1] 30/06/2023 Image Augmentation

How to Make Deep Models Strong and Robust with Data Augmentation

Refer to:

• https://bkai.ai/wp-content/uploads/2023/06/02_07-HUST-NAVER-AIs-Research_-Towards-Strong-and-Robust-Deep-Models.pdf
• https://github.com/ClovaAI/CutMixPyTorch

Problems of training Deep Learning models

• Robustness

• Generalization

Data Augmentation: Simple but very effective solution

• Random Crop => handle scale and viewpoint change cases

• Color jttering/lighting => handle color change cases

• Random erasing/Cutout => handle occlusion cases

• Mixup => handle uncertain images

• CutMix => handle both occlusion and uncertain cases

Unlike Cutout, CutMix uses full image region
Unlike Mixup, CutMix makes realistic local image patches
CutMix is simple: only 20 lines of PyTorch code

Visualize CutMix effectiveness by heatmap

• Mixup, Cutout and CutMix images

• Heatmap results

Experiments with CutMix

• For Image Classification

• For Object Detection and Image Captioning

• Improved robustness performance