Video-based person re-identification with scene and person attributes

Person re-identification (Re-ID) is an essential computer vision task retrieving a person of interest across multiple non-overlapping cameras. In recent years, video-based person Re-ID research has become more and more popular. Compared with image-based person Re-ID, it can obtain more feature information from multiple frames such as temporal information. However, video-based person Re-ID still faces challenges such as occlusion, multiple people and target changes. Given the above issues, a network integrating person attributes feature and scene attributes feature with person feature is proposed to assist person Re-ID. In our method, the feature of person attributes and scene attributes is re-weighted, making it possible to make full use of the person attribute feature when it is difficult to extract the feature of the person in some problematic cases. Moreover, a strip pooling operation is applied to the person Re-ID network. The horizontal and vertical contextual information is extracted separately through the strip pooling operation, leading to an increased receptive field and improved the person Re-ID accuracy. Extensive experiments on MARS and DukeMTMC-VID datasets show that the proposed methods achieve competitive results with state-of-art methods.