// the find
facebookresearch/maskrcnn-benchmark
Fast, modular reference implementation of Instance Segmentation and Object Detection algorithms in PyTorch.
A PyTorch implementation of Faster R-CNN and Mask R-CNN from Facebook AI Research, built around a BoxList/BoxCoder abstraction that several follow-on papers (FCOS, RetinaMask) forked and built on top of. The README says it plainly: this is deprecated in favor of detectron2, so it's useful mainly if you're reading older research code or maintaining something that already depends on it.
The BoxList abstraction in structures/bounding_box.py is a genuinely clean way to carry boxes, labels, and masks together through transforms without juggling parallel arrays - worth reading even if you don't use the repo. Config-driven model assembly (configs/*.yaml + paths_catalog.py) made it easy to swap backbones and heads without touching training code, which is why several papers forked it directly instead of starting from scratch. The custom CUDA kernels in csrc/ (ROIAlign, NMS, deformable conv) were legitimately faster and lighter on GPU memory than Detectron's Caffe2 ops at the time.
Pinned to PyTorch 1.0 - it won't build against any current PyTorch without surgery on the custom C++/CUDA extensions, and the setup.py/compile step was already a common source of pain even when the repo was active. Mixed precision training depends on NVIDIA's standalone APEX, which itself is largely superseded by torch.cuda.amp, so that path is stale on top of stale. No maintenance since the deprecation notice went up - the last push in 2023 was likely just metadata/README, not real fixes - so any issue you hit you're debugging alone. If you need Mask R-CNN today, detectron2 or torchvision's own implementation get you the same ideas with none of the build headaches.