Detectron2

Detectron2 is Meta AI Research's PyTorch library for object detection and segmentation, the successor to the original Detectron and maskrcnn-benchmark. It implements object detection, instance segmentation, panoptic segmentation, DensePose human surface estimation, keypoint detection, and rotated bounding boxes, with architectures spanning Faster R-CNN, Mask R-CNN, Cascade R-CNN, PointRend, DeepLab, ViTDet, and MViTv2. The design is deliberately modular: components can be swapped or extended for research projects, and a large model zoo supplies pretrained baselines for quick starts or fine-tuning. Trained models export to TorchScript or Caffe2 formats for production deployment, and the framework trains substantially faster than its predecessors. Released under the Apache 2.0 license, it serves both academic researchers prototyping new detection methods and production teams, including groups inside Meta, shipping vision systems commercially.