pterhoer/FaceImageQuality

SER-FIQ measures face quality through stochastic embedding robustness—leveraging dropout variations in a face recognition network to estimate how well an image will perform in recognition tasks without requiring manual quality labels. Built on ArcFace embeddings with MXNet, it achieves cross-database generalization by assessing embedding stability rather than training a separate quality model. The approach adds minimal computational overhead (~10% to standard embedding generation) and integrates directly into existing face recognition pipelines, while also documenting demographic bias patterns inherent in quality-recognition coupling.

577 stars. No commits in the last 6 months.