Masking for better discovery: Weakly supervised complementary body regions mining for person re-identification.

Person re-identification still facing several challenges related to many factors such as complex poses, occlusion, misalignment and bad detection. Recent works in the literature focus on extracting local information from the human body but most of them rely on full supervision during training. We propose in this paper a new end-to-end trainable neural network, named Attention Dropping Network (ADN), for diverse rich visual cues discovery without extra human semantic parsing. ADN aims to find fine-grained local information to address the shared person re-identification challenges. Concretely, our network consists of two branches. The Attention Global Branch learns pixel-level local regions based on a fine-grained attention mechanism while the Feature Dropping Branch learns additional missed features in a weakly supervised manner. The fine-grained attention mechanism allows our model to be robust to complex pose variations and to avoid the redundant backgrounds. Extensive experiments over three benchmark datasets (Market-1501, DukeMTMC-reID, and CUHK03) demonstrate the effectiveness and robustness of the proposed network in handling the problems of complex poses, misalignment and occlusions. • Weakly supervised data augmentation based network for fine-grained person re-identification. • Modeling the spatial relationship of a fine-grained region for person re-identification. • Discovering more complementary feature representation in a weakly supervised fashion. [ABSTRACT FROM AUTHOR]