Generalized person re-identification

Abstract

Recently, domain generalized person re-identification(re-ID) has been a hot topic in computer vision research. In recent years, the performance of domain generalized person re-ID has improved significantly. However, these methods usually require large computing resources, including large graphic card’s memory, CPU memory and computational power, which is not practical to the real world scenario. This project present a loss function to replace the traditional one, and is computational resource friendly. This project also uses some existing method to improve the training process, which allow large batch size to be fitted into a single GPU. Through these methods, researcher can train a neural network with less computational resources, achieving the similar performance as training on a larger one.