Semi-supervised evaluation of face recognition in videos

Abstract

With the deployment of cameras for security control or identity verification, video streams become a common input for people tracking, behavior analysis or face recognition. Various methods have been proposed for these purposes, but their evaluation remains a core topic today. Indeed, due to the among of data in video streams, manual annotations limit the ground truth information that can be generated in reasonable time. Moreover, to validate an algorithm, various video streams should be used for the evaluation, to cover the variety of existing scenarios. The manual labeling of such a mass of data represents a considerable task.To avoid this issue, we propose a semi-supervised method to evaluate face recognition performances in videos and easily validate automatic labeling by biometric comparison. Instead of using manual labeling of characters in the videos, the only input needed is the cast of the show/movie, and face images associated to these characters. Comparison between different face coding and comparison algorithms can then easily be lead on a large set of videos. We illustrate our workflow on two series: Prison Break (77 hours of videos for the whole TV series) and the first season of Camera Cafe (10 hours), showing the interest of the approach, both for evaluation and for automatic labeling of people in massive sets of videos.