AutoAugment for Detection Implementation with Pytorch

• Unofficial implementation of the paper Learning Data Augmentation Strategies for Object Detection

0. Develop Environment

Docker Image
- pytorch/pytorch:1.8.1-cuda11.1-cudnn8-devel

• Using Single GPU

1. Implementation Details

• augmentation.py : augmentation class with probability included
• dataset.py : COCO pytorch dataset
• functional.py : augmentation functions for augmentation class
• policy.py : augmentation policy v0, v1, v2, v3, vtest
• Visualize - Bounding Box Geometric Augmentation.ipynb : experiments of bounding box geometric augmentation
• Visualize - Color Augmentation.ipynb : experiments of color augmentation
• Visualize - Geometric Augmentation.ipynb : experiments of geometric augmentation
• Visualize - Magnitude Check.ipynb : experiments for checking Magnitude is right
• Visualize - Other Augmentation.ipynb : experiments of left augmentation
• Visualize - Policy.ipynb : experiments of policy
• Details
  • range are different so just followed the official code not the paper
  • some of the range are fixed cause of mismatch with magnitude
    • range 0.1 ~ 1.9 for color operation (Color, Contrast, Brightness, Sharpness)
    • but 1 is the default (original image)
    • so in this repo, I code like below
      • instead using 0.1 ~ 1.9, use 0 ~ 0.9 with random change (0.5 probability)
      • e.g.) 0.9 was chosen randomly minus the value (0.9 or -0.9) and add with 1 (1.9 or 0.1)
  • do not use numpy nor opencv for speed and preventing version crashes
  • similar design pattern following torchvision transforms code
  • some of the codes can be improved but not considered in this repo (e.g. TranslateX_Only_BBoxes - translate considering bbox size not fixed pixel)

2. Results

2.1. Color Augmentation

2.2. Geometric Augmentation

2.3. Bounding Box Augmentation

2.4. Other Augmentation

2.5. Policy

2.6. Magnitude

3. Reference

• Learning Data Augmentation Strategies for Object Detection [paper] [official code]