A progressive method of face clustering for mobile phone applications

Abstract

We present a progressive method of applying easy-to-hard grouping technique that applies increasingly sophisticated feature descriptors and classifiers on reducing number of image samples from each of the iteratively generated clusters. The primary goal of the proposed approach is to design a cost effective face clustering method to deploy on low-power devices like Mobile phones, while handling various face related real-world challenges like head-pose variations, expressions, make-up, etc. Initially, the proposed method applies K-Means technique with relatively large K value on LBP features of the input faces to create high precision clusters, which may have low recall rates (faces of a single person falls into multiple clusters). The multiple clusters of same individuals generated are then progressively merged by applying sophisticated features like Gabor filters, Gabor jets and sub-space modelling strategies. The method considers reduced number of non-redundant faces from each of the intermediate clusters to the higher stages of clustering to achieve a compute efficient solution that is suitable for mobile devices. Our experiments on the standard databases like YouTube Celebrities, Multi-Pie, Extended Yale-B, CK+, MindReading and internally collected images/videos from mobile phone albums demonstrate the effectiveness of proposed approach as compared to state-of-the-art methods as well as an existing solution on a flagship mobile phone.