Automatic selection of regions of interest in a video by a depth-color image matting

The automatic detection of regions of interest in a video is fundamental for a fast generation of many ground truth images. In this paper we introduce a new solution for selecting regions of interest based on an automatic image matting method. Image matting is a set of techniques designed to obtain a precise separation of background and foreground in image or video sequences. Basically all the matting approaches need a direct human interaction, there are only few total automatic solutions. To achieve this goal we combine two different video streams: the color one and the depth one. In particular, we use an automatic depth based segmentation to substitute the human input in the Soft Scissors, one of the most precise matting algorithm. The overall efficiency is achieved using the Nvidia CUDA architecture to execute the most computational intensive sections of algorithm. The result of the matting can be used as a ground truth for successive elaborations.