The method for defocusing selfie taken by mobile frontal camera using burst shot

Abstract

In this paper, we present a novel human segmentation technique that automatically detects upper-body region in the selfie. To detect and segment upper-body without user interactions, we develop an initial tri-map by combining face detection results and upper-body shape prior in the selfie. Moreover, we employ motion vectors between two images captured in a short time interval to deal with various human poses and cluttered backgrounds. From motion vectors, we estimate the implicit depth layers without auxiliary hardware or time-consuming algorithms. By integrating information from the face detector, shape prior, and motion vectors, we detect and segment the human upper-body accurately. We also implement the proposed algorithm on the mobile phone. In the extensive experiments on selfie dataset, the proposed method shows competitive results in terms of accuracy/recall and outperforms the previous methods.