3D CONTENT-BASED RETRIEVAL WITH SPINIMAGES J¨ urgen Assfalg, Gianpaolo D’Amico, Alberto Del Bimbo, Pietro Pala Dipartimento Sistemi e Informatica Universit` a di Firenze {assfalg,damico,delbimbo,pala}@dsi.unifi.it ABSTRACT Along with images and videos, 3D models have recently gained increasing attention for a number of reasons: advancements in 3D hardware and software technologies, their ever decreasing prices and increasing availability, affordable 3D authoring tools, and the establishment of open standards for 3D data interchange. The ever increasing availability of 3D models demands for tools supporting their effective and efficient management. Among these tools, those enabling content-based retrieval play a key role. In this paper we present a novel approach to 3D content-based retrieval that is based on spin images. Spin images are used to derive a view-independent description of both database and query objects: a set of spin images is first created for each object; then, a descriptor is evaluated for each spin image in the set; clustering is performed on the set of image-based descriptors of each object to achieve a compact representation of the object, thus allowing for efficient indexing and matching. Experimental results are presented for a test database of about 300 models. These results indicate that spin images can be suc- cessfully exploited for content-based retrieval of 3D objects. 1. INTRODUCTION Content-based retrieval of 3D information is a new challenge for researchers and practitioners. Development of techniques support- ing archival, indexing and retrieval of 3D models is of paramount importance in a variety of application domains. For instance, this is particularly the case in the fields of cultural heritage and his- torical relics, where there is an increasing demand for solutions enabling preservation of relevant artworks (e.g. vases, sculptures, and handicrafts) as well as cataloguing and retrieval by content. Tools supporting retrieval of 3D models are also expected to play a key role in educational programs, either traditional or computer- based. In this paper we address the problems of description and match- ing of 3D objects. Inherent difficulties in this field concern the representation of objects, the spatial transformations that may af- fect them, self-occlusions, and varying levels of detail. We present a solution based on the use of spin images ([4]) to capture the global shape of 3D objects in a view independent manner. Since object description based on spin images entails a huge amount of information, feature extraction and clustering techniques are used to meet the specific storage and efficiency requirements of content-based retrieval. By relying on spin images, we provide for an object-centred description, which is insensitive to rigid trans- formations, and which can leverage achievements in the field of content-based image retrieval. The paper is organised as follows: Sec.2 surveys related work in the field of 3D CBR; Sec.3 introduces object description based on spin images; Sec.4 expounds on feature extraction and clusetring; then, in Sec. 5 experimental results are presented; finally, in Sec.6 conclusions are drawn. 2. RELATED WORK Description and retrieval of 3D objects based on description and retrieval of 2D views has been addressed in [9] and [11]. How- ever, the effectiveness of these solutions is limited to description and retrieval of simple objects. In fact, as complex objects are con- sidered, occlusions prevent to capture distinguishing 3D features using 2D views. Description of 3D surface data for the purpose of recognition or retrieval has been addressed for some time. A few authors have investigated analytical 3D models, but this is not always a viable solution, as there are many limitations in providing parameteri- zations of arbitrary models. In [7] retrieval of 3D objects based on similarity of surface segments is addressed. Surface segments model potential docking sites of molecular structures. The pro- posed approach develops on the approximation error of the sur- face. However, assumptions on the form of the function to be ap- proximated limit applicability of the approach to special contexts. Much attention has been recently devoted to free-form (i.e. polygonal) meshes. While this representation of 3D models poses major hurdles to development and implementation of algorithms, it is indeed the most appealing field of application. The system de- veloped within the Nefertiti project supports retrieval of 3D mod- els based on both geometry and appearance (i.e. colour and tex- ture) [12]. Also Kolonias et al. have used dimensions of the bound- ing box (i.e. its aspect ratios) and a binary voxel-based representa- tion of geometry [8]. They further relied on a third feature, namely a set of paths, outlining the shape (model routes). In [10] a method is proposed to select feature points which relies on the evaluation of Gaussian and median curvature maxima, as well as of torsion maxima on the surface. In [5], Elad et al. use moments (up to the 4-7th order) of surface points as basic features to support retrieval of 3D models. Differently from the case of 2D images, evaluation of moments is not affected by (self-)occlusions. In [1] description and retrieval of 3D objects is accomplished through a combination of warping and projection. This method captures prominent geometric features of 3D objects in a view independent manner. However, it can be applied only to objects whose surface defines the boundary of a simply connected 3D re- gion. Moreover, warping may introduce irregular deformation of the object surface before its projection on a 2D map. 0-7803-8603-5/04/$20.00 ©2004 IEEE.