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.
Red boxes are optimized by gradient descent in respect to blue boxes.
- Single box
- Regular boxes on grid
- Irregular boxes on grid
- Random boxes
