Face extraction pipeline that skips the awkward profile shots

What it does

Feed it video with multiple people; it detects faces via MTCNN, tracks them across frames with a Kalman filter (via SORT), and outputs only the “optimal” face per person—meaning frontal poses, not side glances. Extracted faces land in ./facepics, optionally with 5 landmarks drawn on.

The interesting bit

The “optimal face” selection is the practical hook. Rather than dumping every detected frame, it filters for quality—useful when you’re building a clean training set and don’t want CNNs learning from half-profile noise. The README is admirably direct about this being glue code: MTCNN + SORT + some heuristics, wired together for a specific pipeline job.

Key highlights

• MTCNN detection + Kalman/SORT tracking for multi-face persistence across frames
• Automatic filtering: excludes side faces, picks the best frontal shot per tracked identity
• Optional 5-point face landmark overlay via --face_landmarks
• Explicitly pitched for two jobs: CNN training set curation and face recognition backend feeding
• Python 3.5+, TensorFlow, Numba, OpenCV stack

Caveats

• README doesn’t specify how “optimal” is scored—pose angle? detection confidence? both? The logic isn’t visible
• No mention of performance numbers, GPU requirements, or batch processing limits
• Output directory is hardcoded to ./facepics; no config file or CLI path override shown

Verdict

Worth a look if you’re currently hand-curating face training data from video or need a preprocessing step before recognition. Skip it if you need real-time streaming at scale or want fine-grained control over quality thresholds—the README keeps those details to itself.