A pedestrian re-identification model training method and a re-identification method

Abstract

This invention discloses a training method and recognition method for a pedestrian re-identification model. The training method involves obtaining training pairs, including training images and training text; extracting image features from the training images to obtain image feature sequences; extracting text features from the training text to obtain text feature sequences; performing instance alignment learning on the global image representation and the global text representation to obtain instance alignment loss values; performing identity discrimination learning on all local image representations and word-level representations to obtain identity discrimination loss values; and updating the parameters of the initialized pedestrian re-identification model based on the instance alignment loss values ​​and the identity discrimination loss values ​​to obtain a trained pedestrian re-identification model. This training method can provide a pedestrian re-identification model that can help improve the accuracy of pedestrian re-identification. This invention relates to the field of computer vision technology.

Images