Deformable Surface Registration with Extreme Learning Machines

One of the most important open problems in the field of computer-aided design and computer graphics is the task of surface registration for non-isometric cases. One of the approaches of addressing surface registration problem is to find the point-wise correspondence between surfaces using state-of-the-art shape descriptors. This paper introduces an improvement to this approach by means of Extreme Learning Machines. The ELM model is trained to distinguish pairs of corresponding points from non-corresponding ones on the dataset with highly non-isometric distortions between models. The proposed method is compared with original shape descriptors. The results show the increase of accuracy in surface registration task, and also reveal the bottleneck of the state-of-the-art shape descriptors.