Rethinking data collection for person re-identification: active redundancy reduction

Abstract Annotating a large-scale image dataset is very tedious, yet necessary for training person re-identification (re-ID) models. To alleviate such a problem, we present an active redundancy reduction (ARR) framework via training an effective re-ID model with the least labeling efforts. The proposed ARR framework actively selects informative and diverse samples for annotation by estimating their uncertainty and intra-diversity, thus it can significantly reduce the annotation workload. Moreover, we propose a computer-assisted identity recommendation module embedded in the ARR framework to help human annotators to rapidly and accurately label the selected samples. Extensive experiments were carried out on several public re-ID datasets to demonstrate the existence of data redundancy. Experimental results indicate that our method can reduce 57%, 63%, and 49% annotation efforts on the Market1501, MSMT17, and CUHK03, respectively, while maximizing the performance of the re-ID model.