Interesting Receptive Region and Feature Excitation for Partial Person Re-identification.

Partial person ReID tasks have become a research focus recently for it is challenging but significant in practical applications. The major difficulty within partial person ReID is that only incomplete and even noisy person features are available for extraction and matching, which puts forward higher requirement to model robustness. To settle down this problem, our paper proposes a novel IRRFE-ReID model, which includes two major innovations, the interesting receptive region selection module and the feature excitation module. The former module can adaptively select the region of interest from original image while the latter one is applied to distinguish representative person features and weight them during matching. Proven by ablation analysis, these two modules are embeddable and considerably conducive for partial person ReID tasks. Additionally, our IRRFE-ReID model achieves the state-of-the-art performance in two mainstream partial person datasets, PartialReID and PartialiLids, with its Rank1 reaching 85.7% and 74.8% respectively.