Supervoxels for human and clothes in RGB-D image

In order to reconstruct human body and remodel clothes in virtual try on and size-fitting scenario, we preprocess the input point cloud into supervoxels through over-segmentation. Over-segmentation of an image into superpixels is a widely used preprocessing approach in image segmentation. Recently, over-segmentation algorithm, which clusters points with similar feature into a small group, known as supervoxel here, has been used on RGB-D image or the combination of observations from several calibrated RGB-D cameras to reduce occurrence of clusters crossing object boundaries. This method makes use of both color information and the 3D geometric relationships to describe features and segmented regions. We apply this method to RGB-D images of human body and clothes, so that largely reduce the number of regions that must be considered in following remodel process. Experiments were taken on point clouds captured by Kinect. Because the final goal is to predict object boundaries that represent different part of human and clothes, we further compare the recall with SLIC method for boundaries we have marked previously, and give a suggestion on choosing a better seeding radius in our scenario.