3D Registration by Textured Spin-Images

This work is motivated by the desire of exploiting for 3D registration purposes the photometric information current range cameras typically associate to range data. Automatic pairwise 3D registration procedures are two steps procedures with the first step performing an automatic crude estimate of the rigid motion parameters and the second step refining them by the ICP algorithm or some of its variations. Methods for efficiently implementing the first crude automatic estimate are still an open research area. Spin-images are a 3D matching technique very effective in this task. Since spin-images solely exploit geometry information it appears natural to extend their original definition to include texture information. Such an operation can clearly be made in many ways. This work introduces one particular extension of spin-images, called textured spin-images, and demonstrates its performance for 3D registration. It will be seen that textured spin-images enjoy remarkable properties since they can give rigid motion estimates more robust, more precise, more resilient to noise than standard spin-images at a lower computational cost.