Binary patterns for shape description in RGB-D object registration

One of the most important tasks in computer vision systems is the description of the local neighborhood around interest points. In RGB-D images, this task is usually performed by computing descriptors encoded as real-value vectors. The use of binary descriptors, like BRISK, BRIEF or ORB, has proven to be adequate to address this task accurately and efficiently. In this paper, we propose a novel binary pattern that encodes the shape around a given point in a RGB-D image with invariance to rotation and scale. This descriptor is contrasted to well-known state-of-the-art 3D descriptors, namely FPFH, Spin Images, SHOT, PFHRGB and CSHOT in order to test its actual performance on the RGB-D Objects dataset. The experiments performed show that the proposed binary pattern descriptor equals and even outperforms state-of-the-art 3D descriptors with a higher computational efficiency.