AONet: Attentional Occlusion-aware Network for Occluded Person Re-identification
Occluded person Re-identification (Occluded ReID) aims to verify the identity of a pedestrian with occlusion across non-overlapping cameras. Previous works for this task often rely on external tasks, e.g., pose estimation or semantic segmentation, to extract local features over fixed given regions. However, these external models may perform poorly on Occluded ReID, since they themselves are still open problems with no reliable performance guarantee and also not oriented towards ReID tasks to provide discriminative local features. In this paper, we propose an Attentional Occlusion-aware Network (AONet) for Occluded ReID that does not rely on any external tasks. AONet adaptively learns discriminative local features over latent landmark regions by the trainable pattern vectors, and softly weights the summation of landmark-wise similarities based on the occlusion awareness. Also, as there are no ground truth occlusion annotations, we measure the occlusion of landmarks by the awareness scores, when referring to a memorized dictionary storing average landmark features. These awareness scores are then used as a soft weight for training and inferring. Meanwhile, the memorized dictionary is momenta updated according to the landmark features and their awareness scores of each input image. The AONet achieves 53.1% mAP and 66.5% Rank1 on the Occluded-DukeMTMC dataset, significantly outperforming state-of-the-arts without any bells and whistles, and also shows obvious improvements on the holistic datasets Market-1501 and DukeMTMC-reID, as well as partial datasets Partial-REID and Partial-iLIDS. Code and pre-trained models will be released online soon.