This repository implements a CVPR 2019 paper that proposes a novel bounding box regression loss enabling simultaneous learning of bounding box transformation and localization variance. The approach leverages learned variance during non-maximum suppression to merge neighboring boxes, improving localization accuracy. The implementation supports Faster R-CNN with VGG-16 and ResNet-50-FPN Mask R-CNN architectures, achieving significant AP improvements on MS-COCO.