Progressive Feature Alignment for Occluded Person Re-Identification

Abstract

Person Re-Identification aims at searching for the target person across different disjoint cameras. However, in real-world scenes, incomplete imaging of pedestrians occurs from time to time, which causes difficulty in aligning incomplete features. To tackle this problem, we propose a Progressive Feature Alignment (PFA) method that gradually learns coarse-to-fine granularity features. Specifically, the proposed PFA method contains two main modules: 1) The Pseudo Label Generation Module that creates the foreground and human part labels of the input image, and helps network learn to capture the visible human body parts. 2) The Pseudo Label Cleaning Module that addresses semantic inconsistency problem appeared between pseudo labels across different identities. Based on these two modules, PFA learns discriminative global and part features. In training stage, we only utilize the identity groundtruth labels for training, without relying on any external clues and models. In inference stage, we only focus on the shared visible human body parts between two images, achieving progressive feature alignment. We conduct extensive experiments on occluded/partial/holistic ReID datasets, and PFA achieves state-of-the-art and comparable performance, demonstrating the effectiveness of PFA.