API para Detecção e Reconhecimento de Faces: Análise da Viabilidade de Fine-tuning em um Modelo Zero-shot

Abstract

Currently, there are several repositories containing important components for facial recognition. However, to our knowledge, there is no repository that provides all the steps for facial recognition with the current state-of-the- art. To address this issue, we propose the development of an API that integrates facial detection and recognition models. In this work, we trained two models: Retina Face using the Wider Face dataset, achieving a 94% accuracy rate in face detection, and Partial FC zero-shot using MegaFace with a 99.7% accuracy rate on the LFW test. Fine-tuning experiments were conducted on the Partial FC zero-shot model to understand which layers are most affected by fine-tuning. The API, implemented in Flask, successfully achieves the goal of detecting and recognizing faces and is freely available on GitHub (https://github.com/thiagobauce/API_DET_REC.git).