Aggregation attention assisted mutual learning for occluded person re-identification

Occluded person re-identification (ReID) has been becoming a hot research topic in computer vision. Existing works mainly rely on the off-the-shelf human parsing or pose estimation models, which restricts their further development. In this paper, we propose an Aggregation Attention assisted Mutual Learning network (AAML) for the occluded person ReID task, which is completely independent of existing models. Among them, the mutual learning strategy is presented to explore the intrinsic connections between the holistic ReID task and the occluded ReID task through knowledge sharing, which is inspired by the fact that two tasks can benefit from each other. Further, the aggregation attention is proposed to make the network focused on more discriminative regions of the occluded images with assistance of the information embedded in the holistic images. Extensive experiments implemented on several person ReID databases demonstrate the superiority of the AAML method to the state-of-the-arts in both the occluded and holistic person ReID tasks.