jw9730 / ori-giou

GIoU for rotated bounding boxes, vectorized and differentiable

GIoU for oriented bounding boxes

This code is a PyTorch-based, vectorized, and differentiable implementation of Generalized union over intersection (GIoU) loss for rotated bounding boxes.

The base code for axis-aligned box manipulation and optimal matching was taken from DETR by Facebook Research.

For use cases, see test.py.

Demo

Red boxes are optimized by gradient descent in respect to blue boxes.

1. Single box

2. Regular boxes on grid

3. Irregular boxes on grid

4. Random boxes