CN101350035A - Three-dimensional model search method test platform based on content - Google Patents

Abstract

The invention belongs to the computer application field and relates to a three-dimensional model retrieval method test platform on the basis of contents. The invention has the following steps: firstly, standardizing and preprocessing a retrieval three-dimensional model sample which is input and a three-dimensional model in a data base, adopting a method of translation transformation to align the mass center of the model and an origin of coordinates, rotating the model to a uniform angle, carrying out reflection transformation to guarantee the consistency of two three-dimensional models which are mirror images mutually on representation, secondly, extracting the characteristics of an output model which is obtained from the first step, thirdly, carrying out the similarity measurement to the eigenvector which is obtained from the second step, fourthly, displaying a preliminary retrieval result according to a similarity priority sorting method, fifthly, optimizing the retrieval result on the basis of a user feedback method, and sixthly, evaluating the performance of one experimental method through drafting the recall ratio and a recall ratio graph of the method. The method can be applied in the retrieval of various three-dimensional solid models and also can be used to evaluate the performance of a new retrieval method and to select an optimum retrieval method.

Description

Content-based three-dimensional model search method test platform

Technical field

The invention belongs to computer application field, particularly content-based three-dimensional model searching system.Can be applicable to the retrieval of various three-dimensional entity models.

Background technology

In recent years, along with the development that three-dimensional data is obtained technology, methods of three-dimensional graphical modelling and graphic hardware technology, three-dimensional model has become the 4th kind of multimedia data type after sound, image and video, and the quantity of three-dimensional model also is explosive increase.The application of three-dimensional model is also more and more wider, widespread use in the fields such as modelling, virtual reality, 3D recreation, multimedia teaching, video display animation, analog simulation, molecular biology and three-dimensional geographic information system of industrial products.The three-dimensional modeling of high fidelity is wasted time and energy very much, and Fa Da Internet technology provides condition for multiplexing existing three-dimensional model day by day.

The thought of three-dimensional model search originates from three-dimensional model identification, three-dimensional model coupling and CBIR.Content-based three-dimensional model search at first calculates and extracts the feature of three-dimensional model automatically from model data, set up the multidimensional information index of three-dimensional model, in multidimensional feature space, calculate the similarity degree between model to be checked and the object module then, realize browsing and retrieving the three-dimensional modeling data storehouse.Its core is to extract a stack features to represent three-dimensional model, comes the implementation model retrieval by comparing its eigenwert then, and this also is the method for at present main flow.

Compare with two-dimentional multi-medium datas such as images, the content information of three-dimensional model is abundanter, geometric attribute (apex coordinate, normal vector, topological structure etc.) data had both been comprised, surface properties (color, transparency, texture, material etc.) data are also arranged, just because of the diversity of model geometric attribute and surface properties and complicacy, extracting its feature becomes the art technology difficult point, and therefore the accurate retrieval of content-based three-dimensional model also becomes a field difficult problem.

Present newest research results ginseng is shown in Table 1, the 1st patent is based on the three-dimensional model search of two-dimentional sketch, rather than based on the content of three-dimensional model itself, virtually search condition has been carried out stricter control, also be, the two dimension sketch is the condition precedent of carrying out model index, but often the two-dimentional sketch of model is difficult to obtain; Second patent is a system that search function only is provided, and other researchist can not estimate the method for oneself with it.The 3rd patent only is a kind of search method based on different information categories, and the function of the experiment porch of native system can't be provided equally.

In table 1 prior art about the patent of three-dimensional model search

Sequence number	Application number	Patent name
			1	200610114097.4	Method for searching three-dimension model based on two-dimentional sketch
2	200810000155.X	Three-dimensional model searching system
			3	200710195061.8	Method for searching three-dimension model based on axis point set delamination helix information

Summary of the invention

The objective of the invention is in order to solve method for searching three-dimension model manyly, estimate and preferred difficult problem, the test platform of a content-based three-dimensional model search is provided.Can be applicable to the retrieval of various three-dimensional entity models; At dissimilar input models or different model banies, carry out the contrast experiment between the different search methods; Based on this experiment porch, be used for assessing the performance of new search method and select optimum search method.

The objective of the invention is to be achieved through the following technical solutions.

Content-based three-dimensional model search method test platform of the present invention, the concrete steps of its realization are as follows:

Step 1: the retrieval three-dimensional model sample of input and the three-dimensional model in the database are carried out standardization and pre-service; Wherein in order to guarantee that model has translation invariance, the method for employing translation transformation is alignd the barycenter of model with true origin; In order to guarantee rotational invariance, model is rotated to unified angle; In order to guarantee that the convergent-divergent unchangeability need be with scaling of model to a unified yardstick; In order to guarantee that the consistance of two three-dimensional models in expression of mirror image need be carried out reflection transformation each other.

Step 2, the output model to obtaining from step 1 carry out the aspect of model and extract;

Step 3, to the proper vector that step 2 obtains, carry out similarity measurement;

Step 4, show the preliminary search result according to similarity degree priority ordering Faxian;

Step 5, based on the method for user feedback, optimize result for retrieval;

The performance that step 6, the recall ratio by drawing this method and precision ratio figure assess certain experimental technique.

It is distribution of shapes algorithm, thickness histogramming algorithm and complex sphere mapping algorithm that the aspect of model in the described step 2 extracts method for optimizing.

Method for measuring similarity in the described step 3 is preferably: Euclidean distance, Jeffery distance, X2 distance and dozer distance.

Beneficial effect

The test platform of a content-based three-dimensional model search of the present invention provides 12 kinds of effective search methods, can be applicable to the retrieval of various three-dimensional entity models; And, can carry out the contrast experiment between the different search methods at dissimilar input models or different model banies; An also available experiment porch is assessed the performance of new search method and is selected optimum search method.

Description of drawings

Fig. 1 is a process flow diagram of the present invention;

Fig. 2 is a distribution of shapes algorithm flow chart of the present invention;

Fig. 3 is 5 kinds of shape function synoptic diagram of the present invention;

Fig. 4 gets the random point schematic diagram in the triangle of the present invention;

Fig. 5 is a thickness histogramming algorithm process flow diagram of the present invention;

Fig. 6 is a computing method vector weights schematic diagram of the present invention;

Fig. 7 is normal direction, area and the position view of three-dimensional model of the present invention surface dough sheet;

Fig. 8 is a dozer algorithm flow chart of the present invention;

Fig. 9 is a relevant feedback process flow diagram flow chart of the present invention.

Embodiment

The invention will be further described to engage drawings and Examples below.

Embodiment

As shown in Figure 1, the content-based designed retrieval flow of method for searching three-dimension model experiment porch of the present invention is: at first, model to model in the 3 d model library and input inquiry carries out standardization and pre-service, and the final input format of model is the .off form; Extracting the feature of three-dimensional model then, can have 3 kinds for the method for experimental selection, is respectively the distribution of shapes algorithm, thickness histogramming algorithm, complex sphere mapping algorithm.By these three kinds of feature extracting methods, form proper vector and deposit the characteristic vector data storehouse in; Secondly, providing the experimental technique of four kinds of similarity measurements, is respectively Euclidean distance, Jeffery distance, χ ²Distance and dozer distance.Carry out assembled arrangement with three kinds of feature extracting methods, obtain 12 kinds of search methods, select a kind of experiment of retrieving, and show result for retrieval according to similarity degree priority ordering Faxian; Then, according to result for retrieval,, optimize result for retrieval based on the method for user feedback.At last, recall ratio and precision ratio figure by this method method assess its performance, and can finally select optimum search method.

With reference to Fig. 2, the distribution of shapes algorithm flow mainly comprises three committed steps, and promptly determining of shape choice function obtains random point and make up histogram.

(1) selected shape function

People such as Osada have defined following five kinds of shape functions and have measured three-dimensional model, as shown in Figure 3:

● A3 distance: any 3 the formed angle values in expression three-dimensional model surface;

● the D1 distance: expression three-dimensional model center is to the distance of three-dimensional surface any point;

● the D2 distance: expression three-dimensional model surface is the distance of point-to-point transmission arbitrarily;

● D3 distance: any 3 the leg-of-mutton areas of being formed in expression three-dimensional model surface;

● D4 distance: any 4 the tetrahedral volumes of being formed in expression three-dimensional model surface.These five function easy to understand have good noise immunity.

Putting the D that adjusts the distance is tried to achieve by formula 1:

$D=\sqrt{{(x_1-x_2)}^2+{(y_1-y_2)}^2+{(z_1-z_2)}^2}---\left(1\right)$

Because three-dimensional model summit quantity is generally all bigger, the calculated amount of calculating distance between all summits is very big, and is also unrealistic, thus in all summits the n of a picked at random sufficient amount summit, by calculating the distance between these summits, make up some histogram is carried out similarity relatively.The time complexity that calculates the right geometric distance of these points is O (n ²).

(2) random point sampling

Whole model is by tri patch T={t ₁, t ₂..., t _kConstitute, the therefore tri patch number of whole model, and each leg-of-mutton each apex coordinate can conveniently obtain, and concrete steps are as described below.

● read T to internal memory;

● calculate all leg-of-mutton area A (t _i), t _i∈ T deposits leg-of-mutton area in an array, and this array is preserved the area sum of all tri patchs that traveled through in the ergodic process;

● generate one and this array is carried out binary chop, find the call number of the tri patch of corresponding random number, find the probability of a triangle surface just to be proportional to its area like this from zero random number to all triangle area sums of this model surface;

● obtain a three-dimensional data points in triangle inside at random according to the homalographic principle, method is as follows: establishing leg-of-mutton three summits that previous step obtains at random is respectively A, B, C, the random number r between generating two 0 and 1 ₁And r ₂, calculate the position of getting random point P in this triangle inside with formula 2.As shown in Figure 4.

$P=(1-\sqrt{r_1})A+\sqrt{r_1}(1-r_2)B+\sqrt{r_1}{r}_{2}C---\left(2\right)$

As shown in Figure 4, this triangle can be regarded series of parallel as and form in the slice on BC limit, and is shown in dotted line.Suppose to have an array to write down successively in accordance with the order from top to bottom, the slice dough sheet summation of having visited.Because so square being directly proportional of triangle area and the length of side is corresponding r ₁Slice be in the position of dotted line among the figure.Utilize r ₂Specify in the stochastic sampling point P point on this slice.From above process as can be seen, the position of the point that collects is in the inner evenly distribution of triangle.

Above algorithm guaranteed finally to adopt series of points, can be evenly distributed on the three-dimensional model surface according to the homalographic principle.

(3) make up histogram

Make up histogram and need carry out the normalization of histogram yardstick.This paper carries out normalization according to mean distance.After according to the mean value of a large amount of D2 distances the size degree of freedom being carried out normalization, the mean value of the D2 distance that at first calculating samples in a large number obtains, this length is divided into S range unit, calculating drops on the interior D2 of each range unit apart from number, has so just formed one according to the normalized discrete distribution of shapes of mean distance.Carry out normalization according to mean distance, its characteristics have higher robustness when being the distribution of shapes of the different models of comparison.Normalized discrete shape distributes and can utilize three bezier curve etc. to carry out level and smooth match, forms the continuous shape distribution curve.

With reference to Fig. 5, the committed step of thickness histogramming algorithm comprises the selection of shape function, calculates the distance between the random point, makes up the thickness histogram.

(1) selected shape function and all identical with the distribution of shapes algorithm in the method for model surface stochastic sampling does not repeat them here.

(2) geometric distance of calculating point-to-point transmission, the while is according to the weights of the normal vector computational geometry distance of 2 place triangle surfaces.Method is as follows: corresponding sampled point is A and B, and they lay respectively on two triangles.As shown in Figure 6.

The unit vector that note connects A and 2 directions of B is n, and triangulation method is respectively n to unit vector ₁, n ₂Calculate weights W, wherein d is the geometric distance of A, B point-to-point transmission.As shown in Equation 3:

W＝n·n ₁×n·n ₂/d ² (3)

Here the meaning of weights W is: through after the grab sample process, each adopt point can regard as on the little bin that is positioned at a homalographic.Suppose to distribute thick and fast between the triangle 1,2 and connect line segment between 2 that adopt arbitrarily.At first, near the little bin A and the B is carried out projection on perpendicular to the plane of n, like this, near the line segment density the AB has just been carried out normalizing, and this step is equivalent to each little bin be multiply by weights Dot (n, n ₁) and Dot (n, n ₂).Secondly, the spherical angle that the little bin of A and 2 correspondences of B is opened the AB directional ray is proportional to 1/d ²Therefore, to the space each when all 2 line unifications take in, need multiply by weight 1/d again ², this is equivalent to give the weights that are proportional to its covering solid angle size to line segment AB.

(3) make up the thickness histogram

Be similar to the algorithm of distribution of shapes, the size degree of freedom carried out normalization according to the average length of the line segment of a large amount of samplings.The range unit that corresponding line segment AB falls into, the structure form distribution curve need add 1 with the statistic of correspondence, and structure thickness histogram then adds weights W with the statistic of respective distances unit.In addition, for certain sampling number, the area between distribution of shapes curve and the range coordinate axle is fixed, and the thickness histogram but can not guarantee this point then, therefore also need the probability coordinate axis is carried out normalizing, the probability summation that makes corresponding all thickness is a constant.

In Fig. 7 complex sphere mapping method, according to the definition of convex polyhedron, any 2 line segment that connects in this entity must drop in the entity fully.Obviously a lot of common three-dimensional models often are non-convex polyhedrons; definition according to the Gaussian sphere mapping; this dough sheet that usually can cause having identical normal direction diverse location is mapped in same point; in order to address this problem; in the process of mapping; consider to introduce the spatial positional information of dough sheet; Gaussian sphere is redefined and is a kind of complex sphere; be that each point on the sphere is expanded be a complex points; the real part of point is still by the normal direction and the cartographic represenation of area of dough sheet; the imaginary part of point then by the distance expression of dough sheet to initial point, obviously can not have identical normal direction to the equidistant dough sheet of initial point, and mapping is called the complex sphere mapping in this.Because the complex sphere mapping has comprised normal direction, area and the positional information of object surface simultaneously, therefore the object space geometric properties has been described more comprehensively.Concrete computation process is as follows: establishing ABC is arbitrary dough sheet of a grid three-dimensional model, and its apex coordinate is respectively: (x ₁, y ₁, z ₁), (x ₂, y ₂, z ₂), (x ₃, y ₃, z ₃), the barycenter of dough sheet is (x _c, y _c, z _c), the normal direction of dough sheet is n _k, dough sheet is the distance D of dough sheet barycenter to initial point to the distance definition of initial point _k, the area of dough sheet is A _k, then the definition of the mapping point on the unit complex sphere is as shown in Equation 4:

${\mathrm{CSF}}_{{\stackrel{\‾}{n}}_k}=A_k+i{D}_k---\left(4\right)$

The barycenter of preference pattern is a true origin, and is simultaneously indeformable in order to make that mapping has the convergent-divergent of model, at first model carried out normalization, to guarantee D _kConsistance.

With reference to Fig. 1, after feature extraction is finished dealing with, need carry out similarity measurement, the invention provides the method for four kinds of similarity measurements, be described below respectively.

(1) Euclidean distance (Euclidean Distance), as the formula (5):

$D(x,y)=\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(x_i-y_i)}^2}---\left(5\right)$

(2) χ ²Distance defines as shown in Equation 6:

$D(x,y)=\underset{i}{\mathrm{\Σ}}\frac{{(x_i-y_i)}^2}{x_i+y_i}---\left(6\right)$

(3) Jensen-Shannon distance:

The Jensen-Shannon distance is also referred to as Jeffrey distance (Jeffrey Divergence), is the empirical formula improvement to the Kullback-Leibler distance, than the Kullback-Leibler distance, has better mathematical stability.Define as the formula (7):

$D_{\mathrm{JS}}(H,H^{\′})=\underset{m=1}{\overset{M}{\mathrm{\Σ}}}(H_m\log \frac{2H_m}{H_m+H_m^{\′}}+H_m^{\′}\log \frac{{2H}_m^{\′}}{H_m+H_m^{\′}})---\left(7\right)$

Wherein, H and H ' are respectively the histograms that contains M bin, H _mIt is the empirical probit of m the bin of histogram H.Can prove that the Jensen-Shannon distance satisfies relations such as nonnegativity, symmetry and triangle inequality, is a kind of effective distance measure.

With reference to Fig. 8, be the algorithm flow chart of dozer distance.(Earth Mover ' sDistance) is proposed by people such as Yossi Rubner this method, its basic thought is, between two point sets, except measuring their similarity degree with the geometric distance between the point set, the size of the minimum workload that can also will make of the conversion between them is measured, its physical significance is comparatively clear and definite, if the similarity degree height transforms to a point set the required minimum workload of another point set so and also just lacks naturally.

(1) definition

The discrete point set X and the Y of given limited size

X＝{(x ₁，ω ₁)，(x ₂，ω ₂)，…，(x _m，ω _m)}，Y＝{(y ₁，υ ₁)，(y ₂，υ ₂)，…，(y _m，υ _m)}，

Wherein, x _i, y _jThe expression point exists respectively

The position in space; ω _i, υ _j〉=0, the weights of measuring point; M and n provide the size of point set.Notice that X and Y be equal weight not necessarily, if $\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i>\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\υ}}_j,$ Then claim point set X to weigh (heavy), and point set Y is light (light).

Stream (Flow) between definition point set X and the Y is the matrix F=(f of any m * n size _Ij), f _IjRecord x _iAnd y _jBetween the size of the weights that are complementary.The implication that flows is that the weights of point set that will be heavy move to light point set, a process till light point set quilt covers fully.

Satisfy the stream of following four conditions, be called feasible flow (Feasible Flow):

(1)f _ij≥0，i＝1，…，m，j＝1，…n；

(2) $\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{f}_{\mathrm{ij}}\≤{\mathrm{\ω}}_i,$ i＝1，…，m；

$\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{f}_{\mathrm{ij}}\≤{\mathrm{\υ}}_j,$ j＝1，…，n；

(3) $\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{f}_{\mathrm{ij}}\≤\min (\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i,\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{\υ}),$ i＝1，…，m，j＝1，…，n

Given feasible flow F, it is formula (8) that definition is converted into point set Y institute work (Work) with point set X:

$\mathrm{Work}(F,X,Y)=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{f}_{\mathrm{ij}}{d}_{\mathrm{ij}}---\left(8\right)$

d _Ij=d (x _i, y _j) be x _iTo y _jDistance, commonly used is Euclidean distance.

Make F (X, Y) be the set of X to all feasible flows of Y, then the dozer between point set X and point set Y distance (Earth Mover ' s Distance, EMD) definition promptly from point set X is converted into the least work of point set Y with carrying out after the normalization with a light side's weights as the formula (9):

$\mathrm{EMD}(X,Y)=\frac{\mathrm{mi}{n}_{f\∈F(X,Y)}\mathrm{Work}(F,X,Y)}{\min (\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i,\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\υ}}_j)}---\left(9\right).$

Claims (3)

1. content-based three-dimensional model search method test platform is characterized in that the specific implementation step is as follows:

Step 1, to the input retrieval three-dimensional model sample and the three-dimensional model in the database carry out standardization and pre-service

The method of employing translation transformation is alignd the barycenter of model with true origin; Model is rotated to unified angle; With scaling of model to a unified yardstick; In order to guarantee that the consistance of two three-dimensional models in expression of mirror image need be carried out reflection transformation each other;

Step 2, the output model to obtaining from step 1 carry out the aspect of model and extract;

Step 3, to the proper vector that step 2 obtains, carry out similarity measurement;

Step 4, show the preliminary search result according to similarity degree priority ordering Faxian;

Step 5, based on the method for user feedback, optimize result for retrieval;

The performance that step 6, the recall ratio by drawing this method and precision ratio figure assess certain experimental technique.

2. content-based three-dimensional model search method test platform according to claim 1 is characterized in that: the aspect of model extracting method in the step 2 is preferably distribution of shapes algorithm, thickness histogramming algorithm and complex sphere mapping algorithm.

3. content-based three-dimensional model search method test platform according to claim 1 is characterized in that: the method for measuring similarity in the step 3 is preferably Euclidean distance, Jeffery distance, X2 distance and dozer distance.

Images