Characterizing the Utility of Surveillance Video for Person Re-Identification

Surveillance videos have many applications which can now be accomplished with automated systems. However, the performance of an automated system under realistic circumstances usually does not match the performance produced with a benchmark dataset, due to quality degradations and content variations. In this paper, we discuss several quality and non-quality factors that impact algorithm performance, and design a metric that measures the utility of a video for performing a specific task. Having such measurement allows a user to compress and/or select videos prior to implementing video analytics, reducing both computing power and storage. Here, we choose to study one of the most common applications for video analytics-person re-identification-and we call the corresponding quality measurement “identifiability”.