A METHOD FOR CONTENT-BASED SEARCHING OF 3 D MODEL DATABASES

With the development of computer graphics and digitalizing technologies, 3D model databases are becoming ubiquitous. This paper presents a method for content-based searching for similar 3D models in databases. To assess the similarity between 3D models, shape feature information of models must be extracted and compared. We propose a new 3D shape feature extraction algorithm. Experimental results show that the proposed method achieves good retrieval performance with short computation time.


INTRODUCTION
Nowadays, 3D model databases have emerged in many applications, such as mechanical engineering, medical visualization, virtual reality, and computer animation.To facilitate the management of these databases, we need a way to search and retrieve 3D models.Traditional methods for searching multimedia data use attached information, such as textual annotation.Finding a 3D model by textual keywords suffers from the following problems: Text descriptions may be inaccurate, incorrect, ambiguous, or in a different language.Many 3D models may not have attached text annotation.Manually annotating them is a tedious and error-incurring work.
It is hard for users to query by using just text to describe a complex 3D shape.
For these reasons, it is necessary to develop a method for content-based 3D model searching.Content-based searching means that a search system can find 3D models similar to a query model automatically.The search system provides a "query-by-example" query interface.Users can submit an example model as a query, and then the search system returns models according to their shape similarity.
The key problem with content-based 3D model searching is how to extract the shape feature information from 3D models effectively.Shape feature information is usually expressed as vectors (shape feature vectors).The similarity among 3D models can be measured by computing the distance between shape feature vectors under a predefined metric.Figure 1 illustrates the workflow of a search system for 3D model databases.This paper proposes a new shape feature extraction algorithm, which is an enhanced version of the traditional D2 shape distribution.We implemented a demonstration instruction 3D model search system based on the proposed method.Experiments show that the performance of the system is satisfactory.
The rest of the paper is organized as follows: some related work is summarized in Section 2. In Section 3, we discuss the problem of a traditional D2 approach and describe our method in detail.Section 4 presents a demonstration instruction 3D model search system based on our method with experimental results.Finally, conclusions are given in Section 5.

RELATED WORK
In recent years, researchers have proposed many shape feature extraction algorithms.These algorithms can be generally classified into two groups: variant and invariant (Tangelder & Veltkamp, 2004).Because a 3D model may have an arbitrary position, size, or pose in 3D space, when using a variant shape feature, models need to be normalized for translation, scale, and rotation before assessing shape similarity.Studies have found that the traditional methods for translation and scale normalization provide good search results, but methods for rotation normalization are not robust.The PCA (Principal Component Analysis) is a commonly used method for rotation normalization.In the literature (Funkhouser et al., 2003), it is found that some similar 3D models have different principal axes defined by PCA.
An invariant shape feature describes any transformation of a shape in the same way and doesn't need normalization.
Because of the lack of an effective method for rotation normalization, how to extract a great deal of rotation-invariant shape features attracts great interest.Osada et al. (2002) proposed a rotation invariant shape feature extraction algorithm called D2 shape distribution.D2 describes the shape feature of a 3D model by calculating the distribution of distances between random points on the surface.D2 has many good properties, such as quickness in computation and concise storage.Experiments by Osada et al. demonstrate that D2 can effectively discriminate grossly dissimilar models.However, it is possible that dissimilar models may have very similar D2 histograms.To reduce this problem, this paper proposes a new D2 shape distribution that works on 3D models represented as voxel models.

SHAPE FEATURE EXTRACTION
Osada's D2 algorithm works on 3D mesh models.A mesh model defines a 3D object by describing its surface.
There are two main steps in Osada's approach: 1) generating random points on the surface of a 3D mesh, and 2) calculating distances between every pair of points and forming a histogram of these distances (see Figure 2).In a D2 histogram, the horizontal coordinate denotes the distance between points, and the vertical coordinate denotes the calculated frequency of the D2 distance distribution function.For a given distance d and a point set P , the D2 distance distribution function at value d can be expressed by the following equation: 2 { , . .} 2( ) p q P s t p q d D d P where q p − denotes the Euclidean distance between p and q , and • is the cardinality of a set.

Figure 2. D2 shape distribution
Because Osada's D2 approach only uses a simple distance distribution to describe 3D shape, it is possible that dissimilar models have similar D2 histograms.We can see in Figure 3 that a table and a car have very similar D2 histograms.

Figure 3. D2 histograms of a table and a car
To improve the discriminability of D2, we can enrich the distance information by differentiating the inner and outer line segments between a pair of sampling points.This idea is illustrated in Figure 4, where  The DIR of the table indicates there are many low ratios because the table is generally a concave model and many line segments lie outside it.In contrast, the car is rather convex, so there are a large number of line segments inside it, and its DIR is dominated by high ratios.Generally DIR provides additional shape information than provided by D2.We can use the DIR to filter out dissimilar models that D2 is unable to distinguish.The dissimilarity between 3D models A and B can be measured by a weighted sum of D2 and DIR: where 1 w and 2 w are weights for D2 and DIR.In the later of this section, we will discuss how to determine these weights through user feedback.
Suppose that a histogram contains n bins: . The 1 L norm of histograms is computed as the following equation: To facilitate computing the length of inner and outer line segments, we convert 3D models into a voxel model.The voxel model is a type of 3D data that uses volume elements (voxels) to represent an object in discrete 3D space (Kaufman et al., 1993).A voxel is a cubic unit of volume that can be seen as the 3D counterpart of the 2D pixel representing a unit of area.Mesh models just define the surface of objects.Thus it is difficult to determine which part of a line is inside or outside the models.On the contrary, a voxel model is a 3D solid, of which the inner and outer are explicitly defined.It is easy to determine which part of a line is inside or outside on voxel model.For surface-based models in 3D continuous space, such as mesh or B-rep, the computer graphics community has proposed algorithms that can convert them to voxel models very quickly by exploiting the hardware acceleration of display adapters.Figure 6 illustrates a 3D mesh model and its corresponding voxel model.The voxel model also provides a way to accelerate the calculation of DIR.When calculating the length of an inner line segment, one needs to repeatedly calculate whether the line intersects with a voxel.To accelerate the line-voxel intersection calculation, we can employ the octree-based spatial subdivision.Octree (Samet, 1990) is a hierarchical representation that subdivides the full voxel space into octants.When testing the intersection between the line and voxels, the octree structure can reduce the computational complexity significantly by searching through the voxel space hierarchically.
In Equation ( 2), we use a weighted sum to combine two metrics D2 and DIR into a unified metric.To determine the weights w 1 and w 2 , we use the information from Relevance Feedback (RF) to estimate them.RF makes the search process an interaction between the computer and user.For an RF-based search process, the system first retrieves similar models and returns them to the user.Then, the user provides feedback regarding the relevance of some of the retrieval results (the user marks the relevant models in the results and submits them back to the search system).Finally, the system uses the feedback information to improve the performance in the next iteration (Baeza-Yates et al., 1999, Rui et al., 1998).
w 1 and w 2 should respectively reflect the effectiveness of D2 and DIR in retrieval.The more tightly a metric makes the known relevant objects distribute in its feature space, the more effective it is.Suppose that q denotes a query and R is the set of relevant models marked by user in the initial retrieval results.w 1 and w 2 are estimated by the following equations: For a given element ' i r , maxdis is the maximum of distances between ' i r and all other elements in ' R under the metric D. D denotes the metric D2 or DIR (when t=1, D is D2; when t=2, D is DIR).The sum of maxdis over all elements in ' R reflects the tightness of known relevant objects.A smaller sum indicates a tighter set and vice versa.
In our tests, the initial values of w 1 and w 2 were both set to 0.5.Then we use over 200 queries and corresponding feedback sets to estimate heuristically the weights w 1 and w 2. We obtain that 1 w =0.63 and 2 w =0.37.

Figure 1 .
Figure 1.Workflow of a searching system for 3D model database as a supplement to help the D2 algorithm filter out dissimilar models.We denote the distribution of this ratio as DIR.

Figure 4 .Figure 5 .
Figure 4. Inner and outer line segments between two sampling points

L
norm of the histograms of D2 and DIR respectively:

Figure 6 .
Figure 6.A mesh model and the corresponding voxel model

Figure 8 .
Search results of the proposed system (a) results of a query model of human body (b) results of a query model of sword

Table 1
lists the average computation time of the four methods.The computation time of the proposed enhanced D2 method is much less than that of sphere harmonics and is only a little more than that of Osada's D2 approach.

Table 1 .
Average computation time