Cross perspective person ReID aerial and ground

Abstract

This dissertation experiments upon cross-perspective person ReID problems. Firstly, dataset rarity is a common-existing problem for person ReID because of less research focus and the difficulty to conduct. Based on existed works, the dissertation introduces a method for modifying existing datasets to better serve person ReID training process. With help of this method, future research work on person ReID datasets can result with more precise human boundary and straight posture. Secondly, cross-perspective person ReID is a problem which has significant practical value but receives comparable less research focus. This dissertation conducted cross perspective ReID experiments based on a GTA-captured person image database and summarized conclusions and analyzed existing problems for cross-perspective person ReID tasks. The conclusion from experiment result analysis shows similarity with existed person ReID common sense and potential research directions for future study.