Vehicle Re-identification with a Pose-Aware Discriminative Part Learning Model

Abstract

Vehicle re-identification (re-ID) refers to the task of identifying specific vehicles within a large-scale database. Some existing re-ID models predominantly treat images as holistic entities when learning feature representations. Nevertheless, features learned solely from global appearance may compromise their discriminative power when it comes to capturing local details, especially due to similarities in vehicle model or color. And some other studies considered to combine the local features. But they ignored the difference among the same vehicle local parts from different vehicle poses. To mitigate this challenge, we propose the Pose-aware Discriminative Part Deep Model (PDPM), which aims to learn discriminative feature representations specifically for vehicle re-ID. The PDPM is capable of simultaneously learning global and local part features, enhancing its robustness across varying viewpoints and imaging conditions. Additionally, we propose a novel technique for pose-aware discriminative part detection. This module treats vehicles as rigid bodies and identifies unique local parts by analyzing subtle differences among vehicles in similar poses. Our experimental results validate the effectiveness of the proposed PDPM when compared to the state-of-the-art methods which combining global feature and local feature.