An attempt to create the most accurate, reliable, and general vision transformers for facial recognition at scale for high performance, fast, and scalable real world usage for facial recognition tasks.
- Scalable vit from Apple as backbone
- Qlora linear layers for speed
- QK Norm for stability
- Sparse MultiHead Attention
- Conditional inputs of image, text, or image text pairs
- High generalization with low examples 1-2
- Outragreously fast inference
- Scalable architecture with minimal memory consumption
pip install frvit
Here is a table of some popular open source facial recognition datasets with metadata and source links:
| Dataset | Images | Identities | Format | Task | License | Source |
|---|---|---|---|---|---|---|
| Labeled Faces in the Wild (LFW) | 13,233 | 5,749 | JPEG | Face verification | Creative Commons BY 4.0 | http://vis-www.cs.umass.edu/lfw/ |
| YouTube Faces (YTF) | 3,425 | 1,595 | JPEG | Face verification | Creative Commons BY 4.0 | https://www.cs.tau.ac.il/~wolf/ytfaces/ |
| MegaFace | 1 million | 690,572 | JPEG | Face identification | Creative Commons BY 4.0 | http://megaface.cs.washington.edu/ |
| MS-Celeb-1M | 10 million | 100,000 | JPEG | Face identification | Custom | https://www.microsoft.com/en-us/research/project/ms-celeb-1m-challenge-recognizing-one-million-celebrities-real-world/ |
| CASIA WebFace | 494,414 | 10,575 | JPEG | Face verification | Custom | http://www.cbsr.ia.ac.cn/english/CASIA-WebFace-Database.html |
| FaceScrub | 107,818 | 530 | JPEG | Face identification | Custom | http://vintage.winklerbros.net/facescrub.html |
| VGG Face2 | 3.31 million | 9,131 | JPEG | Face verification, identification | Creative Commons BY 4.0 | https://www.robots.ox.ac.uk/~vgg/data/vgg_face2/ |
| UMD Faces | 8,501 | 3,692 | JPEG | Face identification | Custom | https://www.umdfaces.io/ |
| CelebA | 202,599 | 10,177 | JPEG | Face attribute analysis | Creative Commons BY 4.0 | http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html |
MIT