Abstract
One of the most important topics in the research concerning 3D local descriptors is computational efficiency. The state-of-the-art approach addressing this matter consists in using keypoint detectors that effectively limit the number of points for which the descriptors are computed. However, the choice of keypoints is not trivial and might have negative implications, such as the omission of relevant areas. Instead, focusing on the task of single object detection, we propose a keypoint-less approach to attention focusing in which the full scene is processed in a hierarchical manner: weaker, less rejective and faster classification methods are used as heuristics for increasingly robust descriptors, which allows to use more demanding algorithms at the top level of the hierarchy. We have developed a massively-parallel, open source object recognition framework, which we use to explore the proposed method on demanding, realistic indoor scenes, applying the full power available in modern computers.