Study on Unsupervised Instance Segmentation Models for Person Re-Identification

Abstract

Unsupervised instance segmentation for person re-identification is mainly used in challenging cases such as occluded person re-identification and 3D re-identification. Furthermore, unsupervised instance segmentation can be considered as an auxiliary cue, especially useful for long-term person re-identification using multiple cameras and single images. Several instance segmentation models, one-stage and two-stage, were examined in this study. We considered two main families of one-stage instance segmentation models: YOLO-based and SOLO-based and trained the most interesting of them. Several datasets were used for experiments, including the Market1501 dataset, the MSMT17 dataset, the DukeMTMC dataset, the DukeMTMC-reID dataset, the CUHK03 dataset, and the VIPeR dataset. The Mask R-CNN model demonstrated the best accuracy results and the YOLOACT++ model showed the best computational results in terms of instance segmentation. To compare the accuracy results without and with instance segmentation, the BUC model for person re-identification was used as a basis. The experimental results show an increase in Rank-1 accuracy values by an average of 2.7–4.9%.