Labeling tool used to label images for object detection, with assistance from an object detector of your choice by overriding the model_setup and predict function (see example , yolo,detectron,gluon-cv).
detectron2 uses the Resnest fork. By default if no over ride the tool will use torchvisions pretrained resnet50 fasterRCNN.
Saves images by default in pascalVOC format in an annotations folder created by the tool in the images directory. Allows annotations to be saved in COCO format by selecting the appropriate option from the startup menu. if test_train_split enabled, the tool can split the annotations based on a stratified split of classes or a sequential split provided the images have some sort of sequential numbering eg 1,2,3,4/01,02,03/1002,1003,1004...
For adding augmentations, take a look at utils.py, you could use policy containers from google brains BBAug package which uses imgaug,it basically just replicates all images and adds augmentations to them, new images with aug_ prefix will appear in image folder. This is a naive implementation and can cause over fitting if mild to moderate augmentations are used so be aware.
TODO:// Demo images ,dependency list ,usage