Augmentation tool

A simple augmentation tool to load, augment and save 3D depth/normal prediction data. This repo contains 17 augmentation operations, and supports auto-aug searched augmentation policies. .

Usage

python3 -m venv env
source env/bin/activate
pip install -r requirements.txt

python src/augment.py \
    --source-dirs ./data/rgb ./data/depth ./data/normal \
    --output-dir ./outputs \
    --count 10 

# Alternatively, you may just run
sh ./scripts/do_augmentation.sh

Features

• Modularized structure to facilitate configurable pipeline.

• 17 augmentation operations implemented:

   'Rotate': transform.Rotate,
   'ShearX': transform.ShearX,
   'ShearY': transform.ShearY,
   'TranslateX': transform.TranslateX,
   'TranslateY': transform.TranslateY,
   'AutoContrast': transform.AutoContrast,
   'Invert': transform.Invert,
   'Equalize': transform.Equalize,
   'Mirror': transform.Mirror,
   'Flip': transform.Flip,
   'Solarize': transform.Solarize,
   'Posterize': transform.Posterize,
   'Contrast': transform.Contrast,
   'Color': transform.Color,
   'Brightness': transform.Brightness,
   'Sharpness': transform.Sharpness,
   'RandomResizedCrop': transform.RandomResizedCrop
  

• Operations, and their sequence & magnitude ranges can be modified by a standalone config file.

• Support Auto Augmentation searched policies when --pipeline RL_searched specified.

• Load and process batch-wise data. Allow data shuffling before loading when --shuffle-load specified.

• Photo metric distortions, like Contrast, Color, Solarize, can be turned on/off for depth and normal data with --photo-distort-all specified or not. Default setting is to only do photo metric distortions on rgb data and apply geometry distortions across all sources.

References

• Auto Augmentation