Performance evaluation of 3D keypoint detectors for time-of-flight depth data

Abstract

In object recognition techniques, specially feature-based methods, a fundamental step is to extract keypoints which are distinct and considerably interesting in the image. There are many different keypoint detectors already available, each with its own specific use and results vary enormously. It is widely agreed that evaluation of feature detectors is important. To our knowledge there is no comparative study for the performance of keypoint detectors for only depth data from Time-Of-Flight (ToF) camera. As ToF sensors are cheap and extensively used for robotic applications, especially sensors with low sensor noise like Swiss Ranger SR-4k which give only depth data, need arises for this comparative study. A meticulous acquisition of different household object's depth data has been achieved using a Cartesian robot. The pose information from the robot has been used for more accurate evaluation. Different keypoints valid for only depth data are extracted and their repeatability is evaluated. A Comparative study has also been done on standard RGB-D datasets using the new metrics we have defined, to test the correctness of our approach with state of the art approaches which have used RGB-D data.