A GibbsBoost Face Detector and its Application to Facial Occlusion Spotting

Abstract

We propose non-deterministic methods for automatically detecting occlusions of people's faces by superimposed symbols. We trained a face detector using an ensemble-learning algorithm (GibbsBoost) that is based on the sequential Monte Carlo method. We implemented an occlusion detector using a mixture of two discriminant functions that were related to the size of the detected face region and the occluded face area. One realization of this detector achieved a true positive detection rate of 90%. We present experimental results and discuss possibilities for further improvements.