Contour-Guided Person Re-identification.

Feature representation is one of the crucial components in person re-identification(re-ID). Recently, local feature has attracted great attention from the re-ID community, and extra visual cues have been well exploited to guide local feature learning, such as pose cues, semantic parsing and etc . Besides, the latest research demonstrates that general CNN-based deep models have a bias to texture feature in pattern recognition, but ignore shape-based feature, which has been verified as significant for cross-domain invariance. As far as we know, there is little work focusing on shape-based feature on person re-ID. In this paper, we introduce a new data modality, pedestrian contour, into the re-ID community, which to our best knowledge is the first attempt to utilize contour explicitly in deep re-ID models. We hypothesize that, as an alternative of other exploited visual cues, pedestrian contour could guide deep models to learn robust shape-based feature, with build-in prior information. We propose several contour-guided architectures to explicitly use pedestrian contour, including plain ones and multi-scale one. Extensive experiments have validated the effectiveness of our models. Moreover, we transfer the methodology into a powerful part-based model, Part-based Convolutional Baseline(PCB), and boost the model performance, which verifies the promising prospect of contour-guided models to expand as an auxiliary mechanism in re-ID.