文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 工程科技 > 交通运输 > 3d model retrieval
侵权投诉

3d model retrieval

doc.guandang.net

Computer Graphics and Image Processing (CGIP)

Institut für Informatik, Universität LeipzigAugustusplatz 10-11, D-04109 Leipzig

Head: Prof. Dr. Dietmar Saupe

3D Model Retrieval

'HMDQ 9 9UDQL`

April 4, 2001

one chose a different scale, position, rotation, ororientation of an original model, then therepresentation in the canonical coordinate framewould still be the same. Moreover, since objects mayhave different levels-of-detail (e.g., after a meshsimplification to reduce the number of polygons),their normalized representations should be the same asmuch as possible. The normalization step ensures thatmodels can be retrieved regardless of the choices theirauthors have made for their mesh representation.

Feature extraction. The features capture the 3D shapeof the objects. Proposed features range from simplebounding box parameters to complex image-basedrepresentations [2]. Usually, the features are stored asvectors with real-valued components and fixeddimension. There is a tradeoff between the requiredstorage, computational complexity, and the resultingretrieval performance.

Similarity search. The features are designed so thatsimilar 3D-objects are attributed vectors that are closein feature vector space. Using a suitable metric nearestneighbors are computed and ranked. A variablenumber of objects are thus retrieved by listing the topranking items.

Abstract

The topic of this project is 3D-object retrieval in which amodel, a polygonal mesh, serves as a query and similarobjects (by shape) are retrieved from a collection of 3Dobjects. Algorithms proceed first by a normalization step(pose estimation) in which models are transformed into acanonical coordinate frame. Second, feature vectors areextracted and compared with those derived fromnormalized models in the search space. Using a metric inthe feature vector space nearest neighbors are computedand ranked. Objects thus retrieved are displayed forinspection, selection, and processing. For the poseestimation we have introduced a modified Karhunen-Loeve transform that takes into account not only verticesor polygon centroids from the 3D models but all points inthe polygons of the objects. Since there is no theory aboutthe optimal way to describe a 3D-object, a variety offeature vectors is included in our system. The performanceof the feature vectors is evaluated using a type ofprecision/recall test. Our implementation demonstratesand visualizes these tools.Introduction

Objects in databases traditionally have been accessedusing attached information such as textual annotation.Recently, methods for retrieving multimedia documentsusing audio-visual content as a key are developed andstandardized in MPEG-7. Many similarity-based retrievalsystems were designed for still image, audio and video,while only a few techniques for content-based 3D modelretrieval have been reported. In this paper we discuss twotools for 3D object retrieval in which a 3D model given asa triangle mesh serves as a query key and similar objectsare retrieved from a collection of 3D objects. Content-based 3D model retrieval algorithms typically proceed inthree steps:

Normalization (pose estimation). 3D models are given

in arbitrary units of measurement and in unpredictablepositions and orientations in 3D-space. Thenormalization step transforms model into a canonicalcoordinate frame. The goal of this procedure is that if

Canonical coordinate frame

There have been several approaches for the normalizationstep, the most prominent one being the PrincipleComponent Analysis (PCA) that produces an affinetransformation of space, that is also known as theKarhunen-Loeve or Hotelling transform. The transform isdefined by a set of vectors, e.g., the set of vertices of a 3Dmodel. After a translation of the set moving its center ofmass to the origin of the coordinate system a rotation isapplied so that the largest spread of the transformed points(the variance) is along the x-axis. Then a rotation aroundthe x-axis is carried out so that the maximal spread in theyz-plane occurs along the y-axis. Finally, the object isscaled to a certain unit size. Essentially, that is theapproach taken in some proposal (see references in [1]). Aserious problem is that differing sizes of triangles are nottaken into account which may cause widely varyingnormalized coordinate frames for models that are identical

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email