Single-scale robust feature representation for occluded person re-identification

Occluded person re-identification (Re-ID) task has been a long-standing challenge since occlusions inevitably lead to the deficiency of pedestrian information. Most existing methods tackle the challenge by employing auxiliary models, including pose estimation or graph matching models, to learn multi-scale or part-level features. However, the methods heavily rely on the external cues, the performance degrades when the target pedestrian is occluded severely or occluded by another pedestrian. This paper develops a novel Re-ID model single-scale robust feature representation (SRFR) to learn discriminative single-scale features without external cues. Specifically, a light-weight spatial memory module is investigated which takes the advantages of key-value memory network to store occlusion features and utilizes self-attention architecture to get fine-grained features. Furthermore, a camera-constrained triplet loss (CTL) function is exploited to mitigate the negative effects of different pedestrian samples under the same camera on the basis of the triplet loss. Experimental results show the SRFR achieves superior performance on both occluded and holistic datasets, which prove that single-scale features can also work well on mining discriminative features.