Person re-identification with part prediction alignment

Abstract

The key to success of person re-identification(re-id) is extracting the discriminative person features. Various part-level feature extraction methods are proposed to capture local person features for re-id. A prerequisite of part feature extraction is that each part should be well located. We believe that ID predictions in different parts of the same image should be consistent. Instead of using the external dataset and pose estimator for guiding, we propose a re-id model with Part Prediction Alignment (PPA), which aims at aligning the predicted distributions between each part. Due to the global feature and local feature contains different spacial information, we consider that the combination of two sides will further improve the detection effect. Therefore, in this paper we adopt the teacher–student training strategy based on PPA for global–local feature extraction, and the global feature extraction branch as a teacher to guide the training of local feature branch. Experimental results on Market-1501, DukeMTMC-reID and CUHK03 (including CUHK03_Detected and CUHK03_Labeled) datasets confirm the effectiveness of our proposal, we achieve Rank1 with 92.4%, 85.1%, 65.5%, 69.2% on Market-1501, DukeMTMC-reID, CUHK03_Detected and CUHK03_Labeled, respectively.