FaceImageQuality

Why we included this project

Face recognition systems quietly fail on low-quality images, and it is often hard to tell which frames are hurting matching accuracy. SER-FIQ, the method at the heart of this repository, sidesteps the usual need for human or synthetic quality labels by measuring how stable a face's embedding is across random subnetworks of the recognition model itself. A face that produces consistent embeddings is treated as high quality, which makes the approach useful for picking frames in video surveillance, screening enrollment photos, or flagging a face for re-capture before it reaches the matcher. The repo also covers bias in face quality assessment and related work, so you can weigh fairness concerns and see how the method compares with supervised alternatives. It is a research implementation rather than a drop-in service, so plan to adapt the code to your own face model and pipeline.