Deep-learning face recognition, circa 2016 CMU

What it does

This CMU project trains and runs face recognition with deep neural networks, mixing Lua/Torch model code with Python utilities for generating face representations, comparing images, and training classifiers. It also includes evaluation scripts for the Labeled Faces in the Wild dataset and demos that can run a trained classifier on a live webcam stream.

The interesting bit

The tech report explicitly targets mobile applications, but the README focuses on desktop demos like webcam classification and a web interface, leaving the mobile story largely untold inside the repo. That gap makes the repository feel more like a well-equipped research snapshot than a product roadmap.

Key highlights

• Ships with training scripts, pre-trained Torch models, and a Python inference library.
• Includes batch-represent for generating face embeddings and LFW accuracy evaluation scripts.
• Provides ready-made demos: compare.py for image pairs, vis-outputs.lua for network visualization, classifier.py for training, and classifier_webcam.py for live webcam classification.
• Bundles a 68-point face landmark detector from dlib.
• Apache 2.0 licensed, with third-party components clearly marked.

Caveats

• The README is sparse on technical details; it does not list model architecture, accuracy benchmarks, or hardware requirements, so you will need to dig through the code or the 2016 tech report for specifics.
• The release is pinned at 0.2.1 and the citation is dated 2016, so this is legacy code in a field that moves fast.
• The codebase straddles Lua/Torch and Python, which may feel like an archaeological layer cake if you are used to current deep-learning frameworks.

Verdict

Worth a look if you are studying the history of open-source face recognition or need a fully packaged 2016-era pipeline. Skip it if you want contemporary performance numbers and modern framework support.