What it does Detectron is Facebook AI Research’s Python toolkit for object detection research, bundling implementations of Mask R-CNN, RetinaNet, Faster R-CNN, and several other algorithms with ResNet, ResNeXt, and VGG backbones. It was built on Caffe2 and designed for rapid prototyping of novel detection ideas.

The interesting bit The README itself is a deprecation notice: FAIR directs everyone to Detectron2, a “ground-up rewrite” in PyTorch. This repo is essentially a 2018 time capsule of influential computer vision research—Mask R-CNN won the Marr Prize at ICCV 2017, and the codebase enabled follow-up work like DensePose and Group Normalization.

Key highlights

• Ships reference implementations of six major detection architectures, including award-winning Mask R-CNN and RetinaNet
• Modular backbone design: ResNet/ResNeXt/FPN/VGG16 supported, with hooks for adding more
• Includes a Model Zoo with pre-trained weights and baseline results
• Apache 2.0 licensed; historically significant citation record (Girshick, He, Dollár, et al.)
• Last notable update was April 2018 (Group Normalization support)

Caveats

• Deprecated. The project explicitly tells users to migrate to Detectron2
• Built on Caffe2, which itself has been absorbed into PyTorch; installation friction is likely higher than modern alternatives
• No active development; issues and PRs appear to be community-only maintenance

Verdict Worth studying if you’re tracing the lineage of modern object detection or reproducing classic FAIR papers. For any new project, use Detectron2 or another current framework instead.