What it does Deep Alignment Network (DAN) is a convolutional neural network that locates facial landmarks—think finding the corners of eyes, nose tip, jawline. The reference implementation reproduces the CVPR 2017 workshop paper, including evaluation scripts for standard benchmarks (300W, LFPW, HELEN, AFW, IBUG, Menpo) and a webcam demo.

The interesting bit The architecture itself isn’t the story here; it’s the archaeological layer cake. The code requires Python 2.7, Theano 0.9.0, and Lasagne 0.2—a stack that was already fading when the paper dropped and is now effectively extinct. The README calmly notes that two TensorFlow reimplementations exist, written by “other GitHub users,” like passing a note in class.

Key highlights

• Pre-trained models available for download (Dropbox and Google Drive links in README)
• CameraDemo.py for live webcam face tracking
• Evaluation scripts compute AUC@0.08 and failure rates with configurable normalization (inter-pupil, inter-ocular, bounding box diagonal)
• Two community TensorFlow ports exist; this repo remains the Lasagne/Theano original
• MIT-licensed code, but models carry 300-W dataset’s research-only restriction

Caveats

• Dependencies are frozen in 2017: Python 2.7, Theano 0.9.0, Lasagne 0.2
• Model download links are external cloud storage (Dropbox/Google Drive), not releases or Git LFS
• No indication of maintenance since publication; the framework dependencies are deprecated

Verdict Grab this if you’re reproducing the 2017 paper exactly or studying historical face-alignment methods. Skip it for production work; use one of the TensorFlow reimplementations or modern alternatives instead.