paragonadey/facial-keypoint-detection

Facial Keypoint Detection

Upload a photo with one or more faces and this app will detect each face (Haar cascade) and predict 68 facial keypoints per face (eyes, eyebrows, nose, mouth, jaw) with a CNN trained from scratch in PyTorch.

• —Mean pixel error on a held-out, labeled test set of 770 images: ~8.5px
• —Full write-up, source code, tests, and an honest limitations section (including a BatchNorm bug found and fixed during development) live in the GitHub repo.

Mouth and jaw keypoints are visibly less precise than eyes/eyebrows/nose on real-world photos — the training data is lower-resolution and more uniformly framed than typical photos, so this is a known, documented limitation rather than a surprise.