Towards effective person search with deep learning: A survey from systematic perspective

Abstract

Person search detects and retrieves simultaneously a query person across uncropped scene images captured by multiple non-overlapping cameras. In light of the deep learning advancement, person search has emerged as a promising research direction that demonstrates great potential for real-world applications. This paper presents a systematic survey of deep learning methods for person search. Different from existing categorizations, we propose a new taxonomy that dissects person search models into four major components i.e., proposal prediction, feature representation learning, training objectives, and ranking optimization. The most representative works in each component are summarized with highlighted contributions to this field. An in-depth analysis is provided upon evaluation performances of state-of-the-art person search models together with a summary of benchmark datasets. Despite that significant progress has been made to date, practical and extendable person search remains an open task. We conclude with discussions on those under-explored yet challenging datasets and learning mechanisms for real-world demands to inspire future research directions.