The 2016 face detector that launched a thousand ports

What it does

This is the reference code for MTCNN, a 2016 method that finds faces in images and performs facial alignment using cascaded convolutional networks. It treats detection and alignment as a single joint optimization rather than two separate chores. The repository provides the MATLAB and Caffe implementation described in the IEEE Signal Processing Letters paper and a follow-up ICCV 2017 work.

The interesting bit

The authors treat the repo less as a product and more as a reference archive: they dutifully link to C++, PyTorch, MXNet, and Caffe rewrites, effectively letting the community maintain their modernization. That honesty is refreshing in academic code releases.

Key highlights

• Original implementation of the IEEE Signal Processing Letters (2016) and ICCV (2017) papers
• Built for MATLAB 2014b+ with a Caffe backend and Piotr Dollar’s toolbox
• MIT licensed
• Authors curate a list of third-party ports (C++, Python/MXNet, Python/Caffe, Python/PyTorch)
• Points to Center Face if you need face recognition rather than just detection and alignment

Caveats

• The dependency stack—MATLAB, Caffe, and an external toolbox—reflects its 2016 origins; the README itself nudges you toward newer reimplementations
• No benchmarks, API docs, or usage examples in the README beyond two result images
• GPU support is optional (CUDA), but the setup is clearly aimed at research workstations, not casual experimentation

Verdict

Grab this if you are reproducing the original paper or auditing historical computer-vision methods. Skip it if you need a modern, maintained library—the authors themselves link to C++, Python, and PyTorch rewrites.