Multi-view Based Pose Alignment Method for Person Re-identification

摘要

This paper proposes a Multi-View based Pose Alignment (MVPA) method for person re-identification (re-id). Most recent methods solve re-id as a matching process based on single image. However, when poses vary or viewpoints change, the performance seriously deteriorates. This paper aims to learn a representation insensitive to view and pose. Specifically, we establish a set of Multi-view based Person Pose Templates (MPPT) and propose a Pose-Guided Person image Generation (iPG~2) model to synthesize multi-view and uniform-pose based images. The representation learned from multi-view images can significantly enhances the accuracy of re-id. We evaluate our method on two popular datasets, i.e., Market-1501 and DukeMTMC-reID. The results show that our framework promotes the performance of re-id a lot and surpass other methods.