This repository provides reference implementations of state-of-the-art instance segmentation and object detection algorithms in PyTorch. It includes optimized RPN, Faster R-CNN, and Mask R-CNN implementations matching or exceeding Detectron accuracy. The library supports multi-GPU training, mixed precision training on NVIDIA tensor cores, batched inference, and CPU inference. It ships with pre-trained models for various Mask R-CNN and Faster R-CNN configurations.