The PyTorch 1.0 detector that beat Detectron, then bowed out

What it does

maskrcnn-benchmark was Facebook Research’s modular training and inference toolkit for object detection and instance segmentation. It implemented RPN, Faster R-CNN, and Mask R-CNN in PyTorch 1.0, shipped with pre-trained models for standard configurations, and offered a helper class to simplify inference pipelines. The project is now deprecated; the README explicitly redirects users to Detectron2.

The interesting bit

At launch it claimed to match or exceed Detectron’s accuracy while training up to twice as fast, and it used roughly 500 MB less GPU memory than mmdetection during training. The codebase also supported mixed-precision training via APEX, CPU-only inference, and batched multi-GPU training—unusual flexibility for a reference implementation of its time.

Key highlights

• Implements RPN, Faster R-CNN, and Mask R-CNN with accuracy matching or exceeding Detectron, per the README
• Claims up to 2× training speed over Detectron and 30% over mmdetection, plus ~500 MB less GPU memory
• Supports mixed-precision training on NVIDIA Tensor Cores, multi-GPU training and inference, and CPU-only inference
• Provides pre-trained models for nearly all reference Mask R-CNN and Faster R-CNN 1× configurations
• Modular BoxList abstraction and dataset hooks intended to make adding custom datasets straightforward

Caveats

• The project is officially deprecated in favor of Detectron2.
• Default configurations assume an 8-GPU setup, so single-GPU training requires manually adjusting batch size, learning rate, and iteration counts to avoid out-of-memory errors.
• Test-time evaluation currently relies on the COCO API, so custom datasets need extra wiring for testing.

Verdict

Worth browsing if you are maintaining legacy PyTorch 1.0 pipelines or studying how modular detection frameworks evolved. Everyone else should follow the README’s own advice and jump straight to Detectron2.