A Comparative Study of Downsampling Techniques for Non-rigid Point Set Registration Using Color

Registration of multiple sets of data into a common coordinate system is an important problem in many areas of computer vision and robotics. Usually a large set of data is involved in the process. Moreover, the sets are in general composed by a large number of 3D points. The input for registration techniques based on point set as inputs make sometimes intractable the process due to time needed to provide a feasible solution to the transformation between data. This problem is harder when the transformation is non-rigid. Correspondence estimation and transformation is usually done for each point in the data set. The size of the input is critical for the processing time and, in consequence, a sampling technique is previously required. In this paper, a comparative study of five sampling techniques is carried out. Specifically, is considered a bilinear sampling, a normal-based, a color-based, a combination of the normal and color-based samplings, and a Growing Neural Gas (GNG) based approach. They have been evaluated to reduce the number of points in the input of two non-rigid registration techniques: the Coherent Point Drift (CPD) and our proposal of a non-rigid registration technique based on CPD that includes color information.