Experimental Evaluation of Quantization Methods in Facial Recognition

Abstract

This work presents an experimental evaluation of the effects of quantizing embedding gene- ration models when applied to face recognition. Two models - Facenet (based on Inception-ResNet) and Transface (based on Vision Transformer) - are compared under different precision formats (FP32 and INT8) and inference backends (Torch and ONNX). Experiments were carried out on the LFW, VGGFace2 and CelebA datasets, assessing Rank-1 accuracy and employing cosine distance as the me- tric similarity metric. The results show that more established models, such as Facenet, are more robust to quantization, maintaining accuracy even in reduced precision formats, where Transface suffers no- ticeable degradation when quantized. Moreover, quantizing the embeddings vectors yielded up to an 80% reduction in storage requirements without significantly impacting performance. These findings underscore the feasibility of quantization as a strategy of optimizing models in resource constrained environments.