Human Body Multiple Parts Parsing for Person Reidentification Based on Xception

Abstract

A mass of information grows explosively in socially networked industries, as extensive data, such as images and texts, is captured by vast sensors. Pedestrians are the main initiators of various activities in socially networked industries, hence, it is very important to quickly obtain relevant information of pedestrians from a large number of images. Person reidentification is an image retrieval technology, which can immediately retrieve target person in abundant images. However, due to the complexity of many important factors especially of changeful poses, occlusion and background clutter, person reidentification still faces extensive challenges. Considering these challenges, robust and distinguishing person representations are hard to be extracted well to identify different people. In this paper, to obtain more discriminative representations, we propose a human body multiple parts parsing (BMPP) architecture, which captures local pixel-level representations from body parts and global representations from whole body simultaneously. Additionally, a straightforward preprocessing method is adopted in this paper to improve the resolution of images in person reidentification benchmarks. To eliminate the negative effects of changeful poses, a simple yet effective representation fusion strategy is used for the original and horizontally flipped images to get final representations. Experimental results indicate that the method proposed in this article attains superior performance to most of state-of-the-art methods on CUHK03 and Market-1501.

Keywords

• Person reidentification
• Semantic parsing
• Global representations
• Local representations