Police-In-The-Loop Person Re-identification

Anonymous

TL;DR: We identify opportunities and challenges of the crime investigation task and propose a specialized HITL framework.

Abstract: Person re-identification (Re-ID) has gained significant interest in the computer vision community. One important application of Re-ID is the analysis of CCTV footages for police investigations. This setting poses a challenge of identifying all the appearances of a given identity on time while also providing an opportunity to utilize feedback from a human investigator. In this paper, we propose a general framework for building human-in-the-loop (HITL) Re-ID systems for criminal investigations. Using this framework, we investigate various HITL re-ranking methods based on relevance feedback and hierarchical clustering. Quantitative evaluations on large scale datasets show significant improvements in Re-ID performance. Code is available at \url{https://}.