Learning global and local features using graph neural networks for person re-identification

Abstract

Person re-identification (re-id) is the task of recognizing an individual across non-overlapping camera views. Some approaches only rely on extracting global appearance features from images and fail to consider people’s local body information (head, foot, body shape), which can be used as complementary. Other techniques combine local and global features but rely on external information such as pedestrian attributes or human pose to locate and align the local regions. This strategy increases the learning difficulty and is not efficient or robust to real-world scenarios. In this paper, we propose an end-to-end deep learning framework to overcome these limitations. Our method combines the global and local feature representations of a pedestrian and captures the body structural information by modeling the spatial relation of patches using graph neural networks. We also represent the relationships between probe–gallery pairs using a graph neural network and propose incorporating a scoring function to mine a correspondence for local regions. Experimental results on several datasets validate the effectiveness of the proposed method.