A Comparative Study on the Quantization of RetinaFace

This study examines the effects of quantization on the RetinaFace model to optimize its size and in- ference speed while preserving accuracy. We apply and evaluate various techniques, including dynamic quantization, Float16, Float16 for latency, Int8, and Int8 for latency, using the WiderFace dataset. Our experiments demonstrate that these methods significantly reduce model size and accelerate inference with minimal accuracy loss. Int8 quantization yields the most substantial reduction in both model size and inference time, while Float16 for latency offers optimal performance for real-time applications. The findings highlight the efficacy of quantization in deploying high-performance face detection models on resource-constrained devices.