Learning to Efficiently Detect Repeatable Interest Points in Depth Data

摘要

Interest point (IP) detection is an important component of many computer vision methods. While there are a number of methods for detecting IPs in RGB images, modalities such as depth images and range scans have seen relatively little work. In this paper, we approach the IP , detection problem from a machine learning viewpoint and formulate it as a regression problem. We learn a regression forest (RF) model that, given an image patch, tells us if there is an IP in the center of the patch. Our RF based method for IP detection allows an easy trade-off between speed and repeatability by adapting the depth and number of trees used for approximating the interest point response maps. The data used for training the RF model is obtained by running state-of-the-art IP detection methods on the depth images. We show further how the IP response map used for training the RF can be specifically designed to increase repeatability by employing 3D models of scenes generated by reconstruction systems such as KinectFusion [1]. Our experiments demonstrate that the use of such data leads to considerably improved IP detection.