Faster R-CNN's Python vacation from MATLAB

What it does

This is a Python port of the official MATLAB implementation of Faster R-CNN, the NIPS 2015 system that merged region proposal networks with object detection. Built atop a fork of Fast R-CNN, it lets you train and test detectors on PASCAL VOC using VGG16 or ZF entirely from Python. The authors have since deprecated it in favor of Detectron.

The interesting bit

The port added an approximate joint training mode that trains the RPN and detection network together instead of alternating, cutting training time by roughly 1.5× on VGG16 while keeping comparable accuracy. The trade-off is a 10% slower test-time speed than the MATLAB original, since some Python layers still execute on the CPU.

Key highlights

• Python reimplementation of the seminal Faster R-CNN paper (Ren et al., NIPS 2015)
• Approximate joint training mode (~1.5× faster training than alternating optimization for VGG16)
• Supports ZF, VGG_CNN_M_1024, and VGG16 architectures
• Models are incompatible with the official MATLAB version due to minor implementation differences
• Explicitly deprecated; authors direct users to Detectron instead

Caveats

• Test-time inference lags the MATLAB original by roughly 10% (220 ms vs. 200 ms per image on VGG16)
• Requires compiling Caffe with Python layer support and can be finicky with recent CUDA/cuDNN versions
• Achieves similar but not identical mAP compared to the official MATLAB code

Verdict

Worth a look if you are studying the evolution of object detection or maintaining legacy Caffe infrastructure. For new work, the authors themselves recommend Detectron.