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

Abstract Vehicle re-identification (re-ID) aims to identify all instances of a specific vehicle in a large-scale database. Existing re-ID models mostly consider images as a whole entity for feature representation learning. However, features learned from global appearance may lose their discriminative abilities for local details, due to similarities in vehicle model or color. To address this issue, we propose a novel Pose-aware Discriminative Part deep Model (PDPM) to learn discriminative feature representation for vehicle re-ID. This PDPM can learn both global and local part features simultaneously, thereby increasing robustness for differing views and imaging conditions. We also introduce a novel pose-aware discriminative part proposal technique, which considers vehicles to be a type of rigid body and can identify distinctive local parts by assessing fine-grain differences among vehicles in similar poses. Experimental results demonstrated the effectiveness of our proposed PDPM in comparison with state-of-the-art methods applied to VehicleID and VeRi-776, two large-scale vehicle re-ID benchmark datasets.