Depth Guided Attention for Person Re-identification

Abstract

Person re-identification is an important video-surveillance task for recognizing people from different non-overlapping camera views. Recently it has gained significant attention upon the introduction of different sensors (i.e. depth cameras) that provide the additional information irrespective of the visual features. Despite recent advances with deep learning models, state-of-the-art re-identification approaches fail to leverage the sensor-based additional information for robust feature representations. Most of these state-of-the-art approaches rely on complex dedicated attention-based architectures for feature fusion and thus become unsuitable for real-time deployment. In this paper, a new deep learning method is proposed for depth guided re-identification. The proposed method takes into account the depth-based additional information in the form of an attention mechanism, unlike state-of-the-art methods of complex architectures. Experimental evaluations on a depth-based benchmark dataset suggest the superiority of our proposed approach over the considered baseline as well as with the state-of-the-art.