except for finer triangle resolution in some parts of themodel. As a solution to this issue we introducedappropriately chosen vertex weights for the PCA [3],while Paquet et al. used centers of gravity of triangles asvectors for the PCA with weights proportional to triangleareas. Such methods improve retrieval results.
For the pose estimation previous we have recentlygeneralized the Karhunen-Loeve transform so that all ofthe (infinitely many) points in the polygons of an objectare equally relevant for the transformation. More detailsabout our ‘‘continuous’’ analysis in the normalization stepcan be found in [1].
In some approaches the pose estimation is based onmoments of solid objects. However, 3D models are notguaranteed to consist of closed surfaces bounding one ormore solids, and it would be a difficult and questionableundertaking to enforce objects to be solids by stitching upsurfaces with boundaries. Therefore, the approach issuitable only for a small class of 3D models.Feature vectors
Since different authors of 3D models have different stylesand use different units, we consider that desired propertiesof feature vectors are invariance with respect totranslation, rotation, reflection, and scaling (provided bythe normalization), robustness with respect to level-of-detail of a model, and changeable dimension.
Our feature vectors (FVs) can be subdivided (by theway of representation) into the following three classes:1. Real vectors in which each component captures acharacteristic of a particular direction (ray-based FV[3]), region of the 3D space (volume-based FV [2]), orregion of a 2D image of a 3D object (depth buffer-based FV [2])
2. Real vectors in which each component corresponds tothe absolute value of a Fourier coefficient. Threedifferent types of Fourier transform (FT) are engaged:one-dimensional FT (silhouette-based FV [2]), two-dimensional FT (applied to depth buffer images), andFT on the 2-sphere (spherical harmonic representation[1]). The Fourier expansions allow embedded multi-resolution feature vectors.
3. Octree structures (voxel-based FV [2]) in which eachcomponent contains information about a correspondingregion (a voxel cell) of the 3D space.In order to measure distance between FVs, we use theL2 norm or a modification of the Hausdorff distance (forvoxel-based FVs).Evaluation
Our 3D-model database used for experiments containedaround 1900 models. The models were mostly gatheredfrom the Internet (3dcafe, Viewpoint, Avalon, etc.). Wehad manually classified models by shape (e.g., cars,planes, bottles, chairs, etc.) and used this classification inso-called precision/recall tests.
Briefly, precision is the proportion of retrieved modelsthat are relevant and recall is the proportion of the relevantmodels actually retrieved. By examining theprecision/recall diagrams for different queries (andclasses) we obtained a measure of the retrievalperformance.
A prototype of an on-line retrieval system provides auser friendly interface with options for selecting the typeof descriptor and matching criterion, browsing thedatabase, and displaying a model in a VRML browser. Aretrieval example generated by this on-line prototype isshown in Figure 1. The models are visualized from the

same direction in the original frame.
Figure 1. Query for a car.
References
[1] Vrani` ' 9
Saupe, D., Richter, J., ‘‘Tools for 3D-object retrieval: Karhunen-Loeve transform and
spherical harmonics’’, IEEE 2001 WorkshopMultimedia Signal Processing, Cannes, France,October 2001, submitted.
[2] Vrani` ' 9
Saupe, D., ‘‘3D Model Retrieval", in:
B. Falcidieno (Ed.): Proceedings of SpringConference on Computer Graphics 2000, ComeniusUniversity Press, Bratislava, Slovakia, pp. 89-93,May 2000.
[3] Heczko, M., Keim, D.A., Saupe, D., Vrani`
D.V.,
‘‘A method for similarity search of 3D objects ’’(inGerman). In: Proceedings of the BTW 2001 (editorsA. Heuer, F. Leymann, D. Priebe), Oldenburg,Germany, Informatik Aktuell, Springer Verlag, pp.384-401, March 2001.
Contact
Dejan V. Vrani`
,vranic@informatik.uni-leipzig.deDietmar Saupe, saupe@informatik.uni-leipzig.de
3D MODEL RETRIEVAL USING GLOBAL AND LOCAL RADIAL DISTANCES Bo Li, Henry Johan School of Computer Engineering Nanyang Technological University Singapore E-...
IEC-Based 3D Model Retrieval System Seiji Okajima and Yoshihiro Okada Abstract. Recently, 3D CG animations have become in great demand for movie and ...
The contour is extracted by the Canny A Probabilistic 3D Model Retrieval System Using Sphere Image 539 edge detection algorithm [17], and is quantized ...
Support Vector Machine active learning for 3D model retrieval_专业资料 暂无评价
Finally, 2D cepstral features extracted from these projection planes will be used for 3D model retrieval. 2.1. 3D Model Alignment First, the pose of ...
Multi-level spherical moments based 3D model retrieval_专业资料。维普资讯 http://www.cqvip.com 10 50 Lie1/hjagUiCEEA0679:5010 ut.JZein nvINC 2 ...
Ontology Framework for Content-Based 3D Model Retrieval_专业资料。24 2 Jura fDonhaUnvriy(nonl gu iestEgo . E)V12d.o.7,No.(01)2 ...
High Efficient Methods of Content-based 3D Model Retrieval_机械/仪表_工程科技_专业资料。CHINESE J0URNAL OF MECHANICAL ENGINEERl1 G ?248 ? Vo1.26,No.2...
Extended cone-curvature based salient points detection and 3D model retrieval_电子/电路_工程科技_专业资料。Multimed Tools Appl (2013) 64:671–693 DOI 10...
暂无评价
热门文档