/rotation_augment

Rotation Augmentation For Yolo / Darknet

Primary LanguagePython

Put images and labels with matching filename into a folder labeled data in git repositories root directory.

This script will translate the yolo bounding box data structure ([center x, center y, width height]) into tf bounding box data structure ([y_min, x_min, y_max, x_max]) in order to do rotations, then will translate back into yolo format.

  1. If you are uncertain if your data is good you can choose option 1 and look through the labeled data visually. If you choose to seperate sorted data when you press spacebar it will move bad data into the badData folder and will move good data into a folder called goodData when you press any other key. Press ESC to escape.

  2. Option 2 will create augmented rotated data and save in the augImages directory. It will only rotate the image if the bounding boxes edges are not within 5% of the images edges. This prevents losing bounding boxes / skewing data.

  3. Option 3 will look through augmented data with bounding boxes drawn. If you choose to seperate sorted data it will move bad data into a folder called badData when you press the spacebar and will move good data into a folder called goodData when you press any other key. Press ESC to escape.

  4. If you are happy with augmented data choose option 4 to move it into your data folder.

  5. Option 5 will create a train text file for ARVP's darknet.

Augment away!