Improving Person Re-identification By Attribute Mining And Reasoning

Because of its practical and research value,pedestrian recognition has attracted the interest of a large number of researchers.Pedestrian attribute analysis plays a vital role in intelligent video surveillance that can facilitate person retrieval.We explores the combination of attributes and pedestrian identity to improve pedestrian re-identification,starting with the simplest multi-task learning method.In the simplest and most direct way,only the correlation between pedestrian attributes and pedestrian identity is considered,but the relevance and relevance are not considered.Further,we find that different attributes have different importance to pedestrian identity information.Based on the considerations above,we propose a novel Graph Convolution Network(GCN)based attribute mining and reasoning(AMR)framework for person re-identification task.At the same time,in order to improve the expressive ability of attribute features,we adopt an Multi-Branch SpatialChannel Attention Ensemble(MBSCAE)module for each attribute feature extraction.With the spatial attention and channel attention embedding,the MBSCAE module can mine the localization and fine-grained information of attributes in a refined fashion.For comprehensive attribute feature learning,we construct a multi-branch variant of spatial and channel attention branches to create a native teacher model for localization ensemble.With the design of MBSCAE module,the AMR net can disentangle attribute features from a global pedestrian feature.Finally,a GCN module is adopted to reason the semantic correspondences between attribute features,which guides a graph based attribute information fusion for person re-identification task.Extensive experimental results based on DukeMTMC-ReID and Market-1501 datasets demonstrate the superiority of our method over state-of-the-art methods from various aspects.What's more,the trained model on Market1501 also shows good generalization ability on CUHK03.