jessengatai/supracut-face-embedding

FaceEmbedding (Core ML) for Supra Cut

A face identity model, converted to Core ML and compiled for the Supra Cut video editor on macOS. Supra Cut uses it while tracking a person: it turns a face crop into a 512-number fingerprint, so the tracker can tell whether a face that comes back into frame is the same person.

Original work

• Network: InceptionResnetV1 from timesler/facenet-pytorch, MIT licence, with its VGGFace2-trained weights.
• Based on FaceNet (Schroff, Kalenichenko, Philbin, "FaceNet: A Unified Embedding for Face Recognition and Clustering", CVPR 2015) and the weights from davidsandberg/facenet.
• Training data: VGGFace2 (Cao et al., 2018).

All credit for the model goes to its authors. This repo only holds a converted copy.

What was done

• The PyTorch InceptionResnetV1 (VGGFace2 weights) was converted to Core ML and compiled to .mlmodelc. Its output matched the PyTorch original at 0.99999 cosine similarity.
• Supra Cut's own (private) model repo, supracut-models, lists this model but keeps no conversion script for it.
• No retraining and no change to the weights.

Files

File	SHA-256
FaceEmbedding.mlmodelc.zip	fe8861d7a8505969f59ef4aa6a9fc79dcfdcbe5180c75e343c58fbbea9960d65

The zip holds FaceEmbedding.mlmodelc. Supra Cut's build checks this hash before it puts the model inside the app.

Licence

MIT, as facenet-pytorch. See the original repo's licence file for the copyright notice.