What it does

This is Facebook AI Research’s training and evaluation code for MultiPathNet, an object detection architecture that extends Fast R-CNN with multiple detection paths. It handles the full pipeline: training on PASCAL VOC or MS COCO, evaluating with standard metrics, and even a demo script that wires in SharpMask proposals to draw bounding boxes on your own images.

The interesting bit

The authors quietly improved on the original Fast R-CNN by fixing ROIPooling bugs, which bumps mAP up about 2 points. The repo also serves as a Rosetta Stone of sorts—pretrained backbones ported from Caffe, TensorFlow, and Torch all coexist here, including a ResNet-18 variant that hits 24.4 mAP on COCO while staying under 90MB.

Key highlights

• Supports AlexNet, VGG, ResNets, Inception-v3, and Network-In-Network backbones
• Multi-GPU training with both data and model parallelism
• Bundles converted proposal files (Selective Search, SharpMask) to sidestep COCO API’s Lua limitations
• Includes pretrained Fast-RCNN models from the original paper, converted to Torch format
• One-command demo pipeline from raw image to annotated output via DeepMask/SharpMask

Caveats

• Requires Linux, NVIDIA GPU compute capability 3.5+, and explicitly conflicts with Anaconda
• Evaluation leans on Python COCO API because the Lua interface can’t handle large proposal files
• Torch-7 itself is effectively archived; this is historical infrastructure, not a modern starting point

Verdict

Worth a look if you’re studying the evolution of two-stage detectors or need to reproduce BMVC 2016 numbers exactly. Skip it if you want something that runs in PyTorch without coaxing a decade-old Lua stack back to life.