LSDNN: Local-Salient Deep Neural Network for Baggage Re-Identification with Material Discerning
Ruibo Chen, Hui Zhang, Chen Li, Yaonan Wang · 2020
Baggage re-identification(re-ID) aims to identify the same baggage across multiple non-overlapping cameras between the pre-check and post-check area, which is preferably challenging. Similar to the person re-ID, the task is also disturbed by the reasons for variations of lighting, occlusion, and motion blurring. Compared to the person re-ID, salient inter-class similarity and and intra-class dissimilarity, which is harder to distinguish. In this paper, we propose a Local-Salient Deep Neural Network (LSDNN), which is composed of a global branch, feature clip branch, and feature split branch supervised by multi-loss. The global branch extracts main appearance features about color and shape. Meanwhile, the other two branches pay more attention to baggage textures and patterns that are the critical characteristics to accomplish accurate classification. In addition, we creatively take material attributes discerning into consideration to alleviate the enormous impact of inter-class similarity and intra-class dissimilarity. Experiments show that the proposed LSDNN achieves 84.7% mAP and 82.4% Rank-1 accuracy on the MVB dataset.