CGSR features: Toward RGB-D image matching using color gradient description of geometrically stable regions

Image local feature extraction and description is one of the basic problems in computer vision and robotics. However it has still many challenges. On the other hand, in recent years, after the appearance of novel sensors like Kinect camera, RGB-D images are easily available. So it is necessary to extend feature extraction and description methods to be applicable on RGB-D images. In this paper we propose a new approach to feature extraction and description for RGB-D images: Color Gradient Description of Geometrically Stable Regions. The proposed method, first finds smooth regions with uniform changes in surface normal vectors. The process in this stage is inspired from MSER algorithm. Each region then is normalized to a fixed size circle and is rotated toward its dominant orientation to make description affine, scale, and rotation invariant. Finally, color gradients log-polar histogram of normalized regions is used for description. Experimental results show that CGSR features have good performance in illumination and viewpoint changes and outperform state of the art techniques such as SURF and BRAND in matching precision and robustness.