Distinctive 3D Surface Entropy Features for Place Recognition Torsten Fiolka 1 , J¨ org St¨ uckler 2 , Dominik A. Klein 3 , Dirk Schulz 1 , and Sven Behnke 2 Abstract—In this paper, we present a variant of SURE, an interest point detector and descriptor for 3D point clouds and depth images and use it for recognizing semantically distinct places in indoor environments. The SURE interest operator selects distinctive points on surfaces by measuring the variation in surface orientation based on surface normals in the local vicinity of a point. Furthermore SURE includes a view-pose- invariant descriptor that captures local surface properties and incorporates colored texture information. In experiments, we compare our approach to a state-of-the-art feature detector in depth images (NARF). Finally, we evaluate the use of SURE features for recognizing places and demonstrate its advantages. Index Terms— surface interest points, local shape-texture de- scriptor, place recognition I. I NTRODUCTION Interest points paired with a descriptor of local image context provide a compact representation of image content. Applications such as place or object recognition require that a detector repeatably finds interest points across images taken from various view poses and under differing lighting conditions. Descriptors, on the other hand, are designed to distinguish well between different shapes and textures. However, one must admit that descriptor distinctiveness depends clearly on the variety of shapes and textures that appear at the selected interest points. Thus, a detector will be preferable, if it selects interest points in various structures and highly expressive regions. We introduced SURE in [1] and propose now a variant and its use for place recognition. SURE uses an entropy-based interest measure to select points on surfaces that exhibit strong local variation in surface orientation. Its descriptor captures local surface curvature properties as well as color and texture cues in case RGB information is available for the points. In experiments, we measure repeatability of our interest points under view pose changes for several scenes and objects and compare our approach with a state-of-the-art detector and descriptor to demonstrate advantages of our approach. We show that SURE is capable of correctly recognizing the semantic label of scenes with a bag-of-words approach. The top row in Fig. 1 gives a short idea how the 1 Fraunhofer Institute for Communication, Information Processing and Ergonomics FKIE, Wachtberg, Germany torsten.fiolka at fkie.fraunhofer.de, dirk.schulz at fkie.fraunhofer.de 2 Autonomous Intelligent Systems, Computer Science Institute VI, Uni- versity of Bonn, Germany stueckler at ais.uni-bonn.de, behnke at cs.uni-bonn.de 3 Intelligent Vision Systems, Computer Science Institute III, University of Bonn, Germany kleind at iai.uni-bonn.de Fig. 1: We detect SURE features in depth images at loca- tions with locally prominent surface curvature. Our interest operator measures the entropy of the distribution of curvature directions at a point in a local neighborhood (top row). The curvature direction (blue arrow) is obtained by the cross product between the estimated surface normal at the point of interest (red arrows) and at neighboring points (green arrows). We propose a descriptor that captures local shape and colored texture at interest points. We recognize places using a bag-of-words approach using SURE features (bottom row). interest point detection works while the bottom row outlines the place recognition application with SURE features. II. RELATED WORK A. Interest Point Detection Feature detection and description has been a very active area of research since decades. Most related to our method, also the entropy measure based on image intensities has been investigated for interest point detection [2], [3], [4]. It has been successfully applied to object recognition [5] due to the high informativeness of maximum entropy regions. However, those methods purely based on intensity image data suffer problems emerging from projective reduction to 2D space [6]. Recently, various methods have been developed to extract interest points from dense, full-view point clouds. Novatnack et al. [7] extract multi-scale geometric interest points from dense point clouds with an associated triangular connectivity mesh. Our approach does not require connec- tivity information given by a mesh. Unnikrishnan et al. [8] derive an interest operator and a scale selection scheme for In Proceedings of 6th European Conference on Mobile Robots (ECMR), Barcelona, Spain, 2013.