Local Structure Detection with Orientation-invariant Radial Configuration Lech Szumilas 1 , Ren´ e Donner 1,2 , Georg Langs 1,2 , Allan Hanbury 1 1 Pattern Recognition and Image Processing Group Vienna University of Technology, Favoritenstr. 9/1832, A-1040 Vienna, Austria 2 Institute for Computer Graphics and Vision Graz University of Technology, Inffeldgasse 16, 8010 Graz, Austria {lech,donner,langs,hanbury}@prip.tuwien.ac.at Abstract Local image descriptors have proved themselves as use- ful tools for many computer vision tasks such as matching points between multiple images of a scene and object recog- nition. Current descriptors, such as SIFT, are designed to match image features with unique local neighborhoods. However, the interest point detectors used with SIFT of- ten fail to select perceptible local structures in the image, and the SIFT descriptor does not directly encode the local neighborhood shape. In this paper we propose a symmetry based interest point detector and radial local structure descriptor which con- sistently captures the majority of basic local image struc- tures and provides a geometrical description of the structure boundaries. This approach concentrates on the extraction of shape properties in image patches, which are an intuitive way to represent local appearance for matching and clas- sification. We explore the specificity and sensitivity of this local descriptor in the context of classification of natural patterns. The implications of the performance comparison with standard approaches like SIFT are discussed. 1. Introduction Local image descriptors have proved themselves to be very useful for the recognition of objects in images. The “bag of key-points” [4] in combination with SIFT descrip- tors [7] are among the most successful techniques [5]. While the SIFT descriptor has been shown to perform well for region matching in transformed images [9], it has the disadvantage that it does not explicitly take the shape of the regions of interest (image patches) into account. This paper presents a new local descriptor, the Orientation-invariant Radial Configuration (ORC) descrip- tor, which extracts shape properties of local image patches and their boundaries at the same time. Figure 1. Example of boundary point detection: giraffe skin on the left and tiger skin on the right. The inner arrows (white) represent a geometrical description of the local structure interior, outer arrows (red) correspond to the local structure exterior. One of the first attempts to capture the shape informa- tion in a local descriptor was proposed by Belongie and Ma- lik [1], called shape context in the form of a log-polar his- togram of the boundary points. Our detector is most closely related to the Intensity-Based Method of Tuytelaars and van Gool [12], which defines interest regions by detecting lu- minance transitions on rays emanating from local intensity extrema. Our approach differs in several ways. The main difference is that instead of fitting an ellipse to a detected region of irregular shape, the ORC descriptor encodes the shape. Furthermore, instead of detecting interest points at local intensity extrema, we use local symmetry extrema — the interest points then tend to appear in the center of salient image structures in the image. We also introduce an ap- proach for detecting multiple pixel value transitions on the rays based on clustering. Finally, the proposed descriptor is able to encode multiple concentric structures in a single descriptor. In this paper we focus on properties of the proposed ORC descriptor and compare ORC with SIFT performance for lo- cal structure matching in different images. In the following two sections we describe the nature of the symmetry based interest points and the construction of the ORC local de-