What it does

faced detects faces in images, video, or webcam streams using two custom TensorFlow networks. It outputs bounding boxes with confidence scores and includes a CLI tool and OpenCV-based annotation utilities. The pitch is simple: deep-learning accuracy without requiring a GPU.

The interesting bit

The author rejected transfer learning from YOLO or MobileNet+SSD because those networks are overbuilt for a single class. Instead, they trained a tiny fully-convolutional network (288×288 → 9×9 grid) for coarse detection, then added a second “corrector” CNN to refine the bounding boxes. It’s an explicit bet that face detection doesn’t need 80-class feature extractors.

Key highlights

• Two-stage architecture: FCNN for detection, standard CNN for bounding-box refinement
• ~5 FPS on a 2015 MacBook Pro i5 CPU; >70 FPS on a Titan XP GPU
• Direct comparison with OpenCV Haar Cascades included in the README, with side-by-side GIFs
• CLI supports images, video files, and live webcam input
• Trained on WIDER FACE dataset

Caveats

• Explicitly marked “Work in progress” and “not a Production ready system”
• Still not on PyPI; install via git+https only
• The 5 FPS CPU claim is on a specific 2015 machine; your mileage will vary

Verdict

Worth a look if you need face detection in CPU-constrained environments and can tolerate prototype-level roughness. Skip it if you need production reliability or already have GPU resources — just use a standard detector.