Learning random forests for segmentation of person in self-portrait photos

We propose a method for person segmentation in self-portrait photographs. We use random forests to learn the joint distribution of shape, texture and color for the hair, skin, clothes and background classes, respectively. Decisions in each tree, based on the distance of the pixel feature to pre-trained exemplars for each feature channel, are selected so that they effectively distinguish each category. Experimental evaluation shows that accuracy of the proposed method is near (less than 7% decline) the state-of-the-art methods which use significantly more complex learning models.