Matting with Sequential Pair Selection Using Graph Transduction

Abstract

We are concerned with the natural image matting problem, where the goal is to estimate the partial opacity of a foreground object so that it can be softly segmented from a background. In sampling-based matting techniques, user interactions are first acquired to provide prior information about foreground and background regions. Samples are then chosen from those interactions to calculate the alpha (opacity) value of every pixel in an image. In this research, we propose a new sampling approach that brings relevant samples to every pixel with an unknown alpha value; this yields accurate alpha maps. We also present two new formulations for objective functions used to assess the suitability of the chosen samples. The evaluation of the proposed method, on the alpha matting online benchmark, shows that its performance is close to the state-of-the-art techniques.