Learning to recognize familiar faces in the real world

We present an incremental and unsupervised face recognition system and evaluate it offline using data which were automatically collected by Mertz, a robotic platform embedded in real human environment. In an eight-day-long experiment, the robot autonomously detects, tracks, and segments face images during spontaneous interactions with over 500 passersby in public spaces and automatically generates a data set of over 100,000 face images. We describe and evaluate a novel face clustering algorithm using these data (without any manual processing) and also on an existing face recognition database. The face clustering algorithm yields good and robust performance despite the extremely noisy data segmented from the realistic and difficult public environment. In an incremental recognition scheme evaluation, the system is correct 74% of the time when it declares ldquoI don´t know this personrdquo and 75.1% of the time when it declares ldquoI know this person, he/she is ...rdquo The latter accuracy improves to 83.8% if the system is allowed some learning curve delay in the beginning.