Description

This repository contains face recognition service: ML scripts & API to use them.

ML Pipeline

ML pipeline contains four main steps:

• face detection to crop faces from images (RetinaFace is used);
• face alignment to descrease the distribution variance;
• feature extraction (InsightFace is used);
• kNN based on cosine distance on top of feature vectors.

ML pipeline is encapsulated into Index class which provides two main methods: add and find. Index can be used in two ways: eyes-only (works better with masked people) and full faces (works better with open faces). See face/ for details.

Examples

Using full face

Using only eyes area

Known issues:

• add CUDA support for faster inference;
• adapt alignment algorithm from InsightFace instead of own implementation to decrease distribution shift;
• nearest neighbours algoritm is full-scan (O(n) complexity, needs to be replaced with approximate NN search, e.g. nmslib).

API / Wrappers

API is written with Tornado framework. End-points:

• /analyze_face - handle post request, that contains image with human face/faces as payload. Return back results of recognition by classifier with extracted embeddings.