An Online Multi-face Tracker for Unconstrained Videos

This paper presents a system for online multi-face tracking in unconstrained videos. Different shooting angles, strong illumination changes, abrupt motion and face pose changes are the main characteristics of these videos. The proposed online multi-face tracking system combines deep convolutional neural network face detection, multiple instances of a tracker based on discriminative scale and space correlation filters, shot change detection, tracking failure detection, tracklet generation, and ResNet-based face identity label assignment. The system is tested on a dataset of YouTube music videos which is characterised by video sequences with great visual differences caused by face appearance variations (changes in pose, size, makeup, and illumination), and/or rapid camera motion. The results of the experiment expressed by MOTA, MOTP and IDS metrics are given and compared with state-of-the-art multi-target trackers.