CN105354593B - A kind of threedimensional model sorting technique based on NMF - Google Patents
A kind of threedimensional model sorting technique based on NMF Download PDFInfo
- Publication number
- CN105354593B CN105354593B CN201510695635.2A CN201510695635A CN105354593B CN 105354593 B CN105354593 B CN 105354593B CN 201510695635 A CN201510695635 A CN 201510695635A CN 105354593 B CN105354593 B CN 105354593B
- Authority
- CN
- China
- Prior art keywords
- threedimensional model
- nmf
- matrix
- collection
- classification
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
Abstract
The threedimensional model sorting technique based on NMF that the invention discloses a kind of, including：Step 1, input threedimensional model collection, feature calculation is carried out to the threedimensional model collection of input and obtains the NMF preliminary classifications of NMF initial inputs matrix and threedimensional model collection as a result, the threedimensional model that wherein threedimensional model is concentrated is the triangle grid model comprising the three-dimensional coordinate of mesh point and the triangle relation of mesh point；Step 2, visualization presentation is carried out to threedimensional model collection, wherein visualization refers to providing visualization interface window for the subsequent interactive operation of user；Step 3, design interacts NMF methods towards threedimensional model collection, according to energy equation, realizes the threedimensional model collection dynamic cataloging based on user's driving, wherein it is respectively to operate division and merging method that can interact NMF methods.
Description
Technical field
The present invention relates to Computer Image Processing and computer graphics techniques field, are based particularly on NMF (nonnegative matrixes
Decompose, Nonnegtive Matrix Factorization) threedimensional model sorting technique.
Background technology
In recent years, with the increasingly raising of powerful modeling software and three-dimensional acquisition equipment availability, digital geometry model money
Source is quickly increasing, and the scale of resources bank, which becomes content that is more huge while including, also to be become more to enrich (such as
All include ten tens of thousands of or even up to a million numbers in the databases such as Trimble/Google 3D warehouse, Turbosquid
Word geometrical model), it rationally reuses these abundant digital geometry resources and contains huge application value.
An important prerequisite for the effective huge profit of resource should be the tissue typing rationalized to Models Sets, so as to
User is better understood from and uses model included in library.Traditional Models Sets tissue classification procedure has mostly used supervision, half
Supervision or unsupervised machine learning method are completed, but these methods only carry out model in library mostly centered on computer
Bottom operation --- taxonomic organization is carried out to Models Sets according to low-level feature, user seldom participates in assorting process, and right
Classification results are few to be got information about, this allows for the histological structure's situation for understanding Models Sets that user is difficult profound, because
This causes effective huge profit of resource with being obstructed.Especially in Models Sets sharp increase, the more numerous overall background of included type
Under, use above-mentioned conventional method that will be increasingly difficult to be competent at related work.
The Models Sets tissue typing mode of more foreground should be made user and computer cooperative design and play respectively respective
Advantage, dynamic cataloging tissue is carried out to Models Sets under the driving of user view, and classification results can be clearly presented
To user, so that user is in the case where intuitively understanding Models Sets, is searched for and obtained needed for it by the heuristic based on browsing
Resource.However such method still suffers from two large problems：Which kind of method is taken to indicate the representation vector and such as of threedimensional model
What shows and interacts classification to carry out visualization to three-dimensional vector.
Invention content
Goal of the invention：It is a kind of based on NMF the technical problem to be solved by the present invention is in view of the deficiencies of the prior art, provide
Threedimensional model sorting technique.
Technical solution：In order to solve the above-mentioned technical problem, the invention discloses a kind of threedimensional model classification side based on NMF
Method, this method carry out semisupervised classification to threedimensional model collection, include the following steps：
Step 1, threedimensional model collection is inputted, carrying out feature calculation to the threedimensional model collection of input obtains NMF initial input squares
Battle array and the NMF preliminary classifications of threedimensional model collection are as a result, the threedimensional model that wherein threedimensional model is concentrated is the three-dimensional for including mesh point
The triangle grid model of the triangle relation of coordinate and mesh point；
Step 2, visualization presentation is carried out to threedimensional model collection, in order to user browse data collection and subsequent modification,
Middle visualization is to provide visualization interface window for the subsequent operation of user；The present invention is based on existing classification annotation information to elder generation
Into t-SNE visualization techniques improved, and browse Models Sets and subsequent operation, design for ease of user and realize
For the visual analyzing prototype system of threedimensional model collection, this is also that first passage provides to visualize and operation is added and is introduced into three
Dimension module is classified；
Step 3, design interacts NMF methods towards threedimensional model collection, according to energy equation, realizes and is based on user's driving
Threedimensional model collection dynamic cataloging, wherein can interact NMF methods be respectively operate division and merging method.
It can understand that the geometry that model includes is believed from the angle of signature analysis using computer in step 1 of the present invention
Breath, and similitude and class inherited in Models Sets class are distinguished using simple effective method, and then complete Models Sets and stablize
Presort, the specific steps are：
Step 1-1 builds bag of words BOW (Bag of words model bag of words) feature of threedimensional model collection,
Obtain NMF initial input matrixes；
Step 1-2 completes the calculating of presorting to threedimensional model collection using NMF, obtains the preliminary classification of threedimensional model collection
As a result.
Step 1-1 includes the following steps：
Step 1-1-1, threedimensional model collection HKS are calculated：Utilize multi-scale diffusion thermonuclear HKS (Heat Kernel
Signature) method carries out feature calculation to each mesh point of input threedimensional model collection, obtains HKS descriptors, to indicate three-dimensional
The local feature information of Models Sets；
Step 1-1-2, vector quantization：The HKS descriptor meters of threedimensional model collection are clustered by k-means (hard clustering algorithm)
Calculation obtains corresponding word list (geometric words)；
Step 1-1-3 builds the probability distribution that threedimensional model concentrates the correspondence word list of each threedimensional model by statistics, obtains
Each threedimensional model bag of words BOW features, while obtaining the input matrix V of NMF using calculated BOW features, i.e., three
Dimension module collection eigenmatrix.
A Visualized Analysis System towards threedimensional model collection is developed in step 2 of the present invention, can be user in vision
It is provided in space and clearly visualizes presentation pattern to show the hoc scenario of Models Sets, and then user and machine can be made more preferable
Cooperation go to obtain more efficiently classification results, specifically include following steps using respective distinctive ability：
Step 2-1 passes through t-SNE methods (Embeddingt points of t-Distributed Stochastic Neighbor
Cloth random neighbor embedded mobile GIS) it completes to turn the bag of words BOW features of each threedimensional model from higher dimensional space to two dimensional surface
It changes；
The visualization interface of threedimensional model collection is added in step 2-2, and t-SNE conversions projection result is shown to complete
It is shown at the visualization of input threedimensional model collection, while the subpoint of each model is corresponded to using NMF preliminary classification results
Label completes visualization display, is used for subsequent operation, and visualization interface window includes auxiliary browser window and visualization area
Domain, wherein the threedimensional model in auxiliary browser window region and the display point of viewable area carry out one-to-one correspondence and show.
Step 3 includes the following steps：
Step 3-1, design interact NMF energy equations towards threedimensional model collection, and the minimum value for seeking the expression formula obtains
Optimal solution is taken, can interact NMF energy equations is：
Formula passes through input parameter V, L, MwIt is best to seek
W, H values.Wherein minW,H≥0It indicates to seek the W under formula value minimizes, H values.
Wherein V is the input matrix for the NMF that step 1-1-3 is obtained, i.e. threedimensional model collection eigenmatrix；W, H are respectively NMF
The required center matrix taken and reference matrix, L are the R-matrix relative to W；MWFor the diagonal matrix of adjustment parameter, matrix
MWFor value on diagonal line between 0~1, α is adjustment parameter, adjusts the weight in equation result convergence direction, range is 0~1
Between；
Wherein the present invention is mainly become by changing L R-matrixes come the classification direction for driving NMF optimal solutions towards user demand
Closely, thus by changing L R-matrixes, the present invention devises two kinds of operations：Classify and merges.
Step 3-2 divides two generic operations dynamic change threedimensional model collection by threedimensional model collection categories combination and classification：
Threedimensional model collection categories combination：After the representative threedimensional model of user's specified three-dimensional Models Sets classification, according to the generation
Classification belonging to table threedimensional model positions its column information in Current central matrix W, during union operation, according to institute
Specified threedimensional model obtains new column information, and the column information not merged in former center matrix W is directly stored in matrix L,
The new cluster centre that the need acquired are merged to threedimensional model again is added in matrix L；
Threedimensional model collection classification divides：User creates completely new classification on the basis of existing classification results, i.e. user passes through
Browsing determines the required classification divided, and selects accordingly to represent the new classification information of obtaining three-dimensional model, when division, original
Increase new cluster centre on basis of classification, NMF is instructed to be calculated, R-matrix L increases on the basis of original center matrix W
Add the column information for representing new cluster centre.
Wherein model selected by division of the invention and union operation, based on the information that visualization interface is provided, user is very
It easily selects that division oneself to be selected finally to obtain optimum by iteration for several times with combined Models Sets.
Advantageous effect：The category of model method advantage of the present invention is：Data mining and machine learning field is the most living
One of research method of jump NMF is introduced into threedimensional model classification field for the first time.In addition, using NMF classification acquired results information,
The present invention improves advanced visualization strategy t-SNE, keeps the visualization display of its classification apparent intuitive, after being more convenient for
Continuous user's operation.Finally, result is presented in conjunction with the visualization of threedimensional model collection, it is proposed that a kind of semi-supervised NMF methods make
It can instruct the modification for having Non-negative Matrix Factorization result, and then realize intuitive, the real-time dynamic to Models Sets classification results
Change.
Description of the drawings
The present invention is done with reference to the accompanying drawings and detailed description and is further illustrated, of the invention is above-mentioned
And/or otherwise advantage will become apparent.
Fig. 1 is broad flow diagram of the present invention.
Fig. 2 is 1 threedimensional model BOF calculation flow charts of embodiment.
Fig. 3 is the result schematic diagram of presorting of embodiment input model collection, is corresponding classification results in each box.
Fig. 4 is that schematic diagram is presented in t-SNE visualizations.
Fig. 5 is that embodiment merges schematic diagram.
Fig. 6 is embodiment into line splitting schematic diagram.
Fig. 7 is final classification result schematic diagram, is corresponding classification results in each box.
Specific implementation mode
The present invention calculates its threedimensional model HKS features first to the threedimensional model collection of input, then by vector quantization come
Threedimensional model BOF is built, the input matrix V of NMF is finally calculated, and it is pre- to complete preliminary NMF to input threedimensional model collection
Classification；Then utilize t-SNE algorithms complete threedimensional model high dimensional feature to low-dimensional feature projection convert and visualize present with
Facilitate user's subsequent operation, finally devises semi-supervised NMF algorithms to be iterated to result of presorting, finally visual
Satisfied threedimensional model classification results are obtained with two generic operations are merged by division under interface.
As shown in Figure 1, the present invention includes the following steps：
Step 1, threedimensional model collection is inputted, carrying out feature calculation to the threedimensional model collection of input obtains NMF initial input squares
Battle array and the NMF preliminary classifications of threedimensional model collection are as a result, the threedimensional model that wherein threedimensional model is concentrated is the three-dimensional for including mesh point
The triangle grid model of the triangle relation of coordinate and mesh point；
Step 2, visualization presentation is carried out to threedimensional model collection, in order to user browse data collection and subsequent modification,
Middle visualization is to provide visualization interface window for the subsequent operation of user；The present invention is based on existing classification annotation information to elder generation
Into t-SNE visualization techniques improved, and browse Models Sets and subsequent operation, design for ease of user and realize
For the visual analyzing prototype system of threedimensional model collection, this is also that first passage provides to visualize and operation is added and is introduced into three
Dimension module is classified；
Step 3, design interacts NMF methods towards threedimensional model collection, and proposes corresponding accounting equation, realizes based on use
The threedimensional model collection dynamic cataloging of family driving, wherein it includes two kinds of operation divisions and merging method that can interact NMF methods, being based on can
It is operated depending on changing interface.
It can understand that the geometry that model includes is believed from the angle of signature analysis using computer in step 1 of the present invention
Breath, and similitude and class inherited in Models Sets class are distinguished using simple effective method, and then complete Models Sets and stablize
Presort, the specific steps are：
Step 1-1 builds bag of words BOW (Bag of words model bag of words) feature of threedimensional model collection,
Obtain NMF initial input matrixes；
Step 1-2 completes the calculating of presorting to threedimensional model collection using NMF, obtains the preliminary classification of threedimensional model collection
As a result.
Step 1-1 includes the following steps：
Step 1-1-1, threedimensional model collection HKS are calculated：Utilize multi-scale diffusion thermonuclear HKS (Heat Kernel
Signature) method carries out feature calculation to each mesh point of input threedimensional model collection, obtains HKS descriptors, to indicate three-dimensional
The local feature information of Models Sets；
Step 1-1-2, vector quantization：The HKS descriptor computations that threedimensional model collection is clustered by k-means obtain accordingly
Word list geometric words；
Step 1-1-3 builds the probability distribution that threedimensional model concentrates the correspondence word list of each threedimensional model by statistics, obtains
Each threedimensional model bag of words BOW features, while obtaining the input matrix V of NMF using calculated BOW features, i.e., three
Dimension module collection eigenmatrix.
A Visualized Analysis System towards threedimensional model collection is developed in step 2 of the present invention, can be user in vision
It is provided in space and clearly visualizes presentation pattern to show the hoc scenario of Models Sets, and then user and machine can be made more preferable
Cooperation go to obtain more efficiently classification results, specifically include following steps using respective distinctive ability：
Step 2-1, being completed by t-SNE (t-Distributed Stochastic Neighbor Embedding) will
Conversion of the bag of words BOW features of each threedimensional model from higher dimensional space to two dimensional surface；
The visualization interface of threedimensional model collection is added in step 2-2, and T-SNE conversions projection result is shown to complete
It is shown at the visualization of input threedimensional model collection, while difference is carried out using NMF preliminary classification results to the subpoint of each model
Label completes visualization display, is used for subsequent operation, and visualization interface window includes auxiliary browser window and visualization area
Domain, wherein the threedimensional model in auxiliary browser window region and the display point of viewable area carry out one-to-one correspondence and show.
Step 3 includes the following steps：
Step 3-1, design interact NMF energy equations towards threedimensional model collection, and the minimum value for seeking the expression formula obtains
Optimal solution is taken, can interact NMF energy equations is：
Formula passes through input parameter V, L, MwIt is best to seek
W, H values.Wherein minW,H≥0It indicates to seek the W under formula value minimizes, H values.
Wherein V is the input matrix for the NMF that step 1-1-3 is obtained, i.e. threedimensional model collection eigenmatrix；W, H are respectively NMF
The required center matrix taken and reference matrix, L are the R-matrix relative to W；MWFor the diagonal matrix of adjustment parameter, matrix
MWValue on diagonal line is between 0~1, and α is adjustment parameter, and range is between 0~1；
Wherein the present invention mainly drives NMF optimal solutions to be approached towards the direction of user by changing L R-matrixes, leads to thus
Modification L R-matrixes are crossed, the present invention devises two kinds of operations：Classify and merges.
Step 3-2 divides two generic operations dynamic change threedimensional model collection by threedimensional model collection categories combination and classification：
Threedimensional model collection categories combination：After the representative threedimensional model of user's specified three-dimensional Models Sets classification, according to the generation
Classification belonging to table threedimensional model positions its column information in Current central matrix W, during union operation, according to institute
Specified threedimensional model obtains new column information, and the column information not merged in former center matrix W is directly stored in matrix L,
The new cluster centre that the need acquired are merged to threedimensional model again is added in matrix L；
Threedimensional model collection classification divides：User creates completely new classification on the basis of existing classification results, i.e. user passes through
Browsing determines the required classification divided, and selects accordingly to represent the new classification information of obtaining three-dimensional model, when division, original
Increase new cluster centre on basis of classification, instructs NMF to carry out completely new calculating, R-matrix L is in original center matrix W bases
Increase the column information for representing new cluster centre on plinth.
Embodiment 1
As shown in Fig. 2, in the present invention：For inputting threedimensional model collection, the three-dimensional mould as shown in (a) in Fig. 2 is inputted first
Then type carries out feature calculation and obtains the threedimensional model HKS thermal maps as shown in (b) in Fig. 2, last to pass through as shown in (c) in Fig. 2
Vector quantization, the distribution calculated on vector space obtain the BOF features of each threedimensional model.
Embodiment 2
Illustrate each step of the present invention below according to embodiment.
Step (1), it is quick, stable according to the isomery Models Sets progress to input to presort, to instruct subsequent visual
Change display and user's operation；
The present embodiment processing Models Sets be size normalizing isomery threedimensional model collection, in order to support it is subsequent visualization and
Interaction NMF operations need the BOF for building model and complete operation of presorting to Models Sets.The present embodiment uses document
BRONSTEIN A.M.,BRONSTEIN M.M.,GUIBASL.J.,OVSJANIKOV M.:Shape google:Geometric
words andexpressions for invariant shape retrieval.ACM Transactions
OnGraphics (TOG) 30,1 (2011), 1.2,4 methods that threedimensional model BOF is built based on HKS proposed.This method
The HKS features of use have multiple dimensioned, the features such as rotational invariance, and the foundation of BOF models can be as the input of NMF matrixes.It is main
It is divided into two processes：BOF is established and NMF presorts.
Step (11), threedimensional model collection BOF are established：
BOF can be good at indicating threedimensional model, therefore for building the non-negative sample matrix V of input, include mainly model
HKS feature calculations, vector quantization and BOF's establishes three steps.
Step (111) threedimensional model HKS is calculated.
The each mesh point for inputting each model of threedimensional model collection is carried out using multi-scale diffusion thermonuclear (HKS) method
Feature calculation, to indicate the local feature information of each model；Assuming that in time t=0, given at three-dimensional model gridding point x
The heat of the heat of one unit, other points of grid is all 0, then allows heat freely to be spread on grid, then kt(x, y) can be with
It is construed to the amount of heat in time t time point y.The k that HKS descriptors are made of different size of tt(x, x) Sequence composition, i.e.,
{kt1(x,x),kt2(x,x),...,km(x, x), } .HKS descriptors not only have Analysis On Multi-scale Features, but also very robust.
Step (112) vector quantization
By above-mentioned calculating, the HKS descriptor features of enormous amount are obtained, it is necessary to carry out vector quantization structure BOF models
Word folder, so clustering the HKS descriptors feature of each model by k-means corresponding word list P=is calculated
{p1,p2,…,pn}；
The foundation of step (113) BOF.
By counting the probability distribution for the word list for building each threedimensional model, the BOW of each model is obtained, this process and picture
BOW to establish process almost the same.
Detailed process is as follows：
After calculating the HKS descriptors feature of all mesh points of model and word list P, begin to calculate the model each
Probability distribution θ (x)={ θ of the point in word list1(x),…,θV(x) }, the present invention is distributed using gaussian probability, meter
It is as follows to calculate formula：
Wherein C (x) is adjustment parameter, in order to be normalized to formula value, i.e., | | θ (x) | |1=1, wherein θ indicate probability
Distribution, p (x) are the HKS descriptors of the point, and pi is a wherein word for the word list that step (112) is calculated.σ is public affairs
Formula variance, value are determined most suitable value by word list.
Step (12) NMF presorts.
The BOF that each model of threedimensional model collection is completed by step (21) is established, and has also just obtained the input matrix of NMF, most
The calculating of presorting to Models Sets is completed using NMF eventually.Here, first, it is necessary to illustrate lower NMF.
The non-negative original sample matrix V that a given size is n × m, namely corresponding to the input matrix of the present invention, it is each
Row correspond to the n dimension nonegative elgenvectors of each sample in m sample, and wherein m determines that n is by word list by sample number size
Word number determine.The purpose of NMF algorithms is exactly to find 2 new matrix Ws and H to carry out approximate original sample matrix V：
In formula, W, H are respectively basic matrix and coefficient matrix, and in the present invention, W, H are the central moment taken required by NMF
Battle array and reference matrix；W is n × r dimensions, and H is that r × m is tieed up, and r is the dimension after dimensionality reduction, i.e. the number of base vector, select it is ensured that
(n+m)r<Nm, it follows that W, H can be considered as just the compressed format of data in V.It can be with by a series of iterative algorithm
W and H are obtained, each row of wherein W all illustrate the cluster of local correlation ingredient to a certain extent.In addition the algorithm is to be based on
What simple iterative process calculated, it is not time-consuming, and also fast convergence rate, result of calculation are stablized, therefore for large-scale data
Classification processing is applicable in very much.
So the Models Sets based on NMF are presorted, it can be stated that following form：
P (features | model)=∑topicP (features | topic) × p (topic | model),
The characteristic probability distribution situation of each model, p (topic │ model) in wherein p (features │ model) library representation
Indicate that model selects theme with certain probability, p (features │ topic) indicates to select with certain probability in this theme
Some feature, therefore the required classification situation for obtaining theme and being final.When by the characteristic probability distribution situation matrix of Models Sets
After being indicated, the above problem can be converted to NMF and solved.It is defeated in the corresponding NMF of wherein p (features │ model)
Enter non-negative sample matrix V, in the corresponding NMF of basic matrix W, p (topic │ model) in the corresponding NMF of p (features │ topic)
Coefficient matrix H, in the present invention, W, H are the center matrix taken required by NMF and reference matrix；.
Based on above-mentioned expression, the BOF vectors of each model of threedimensional model collection are exactly the row of matrix V, namely Models Sets are determined
The non-negative sample matrix V of input.The process of presorting is completed eventually by NMF, as shown in Figure 3.
Step (2) carries out visualization presentation to existing category of model result, in order to user browse data collection and after
Continuous modification；The present invention is based on existing classification annotation information to improve advanced t-SNE visualization techniques, and for ease of
User browses Models Sets and subsequent operation, and design realizes the visual analyzing prototype system for threedimensional model collection；It is main
It is divided into t-SNE visualization displays and shows two parts with auxiliary area model.
Step (21) t-SNE visualization displays
The conversion by the BOW features completion of each threedimensional model from higher dimensional space to two dimensional surface is completed by t-SNE, is completed
The visualization of input threedimensional model collection is shown.It is necessary to first introduce lower t-SNE herein.
Hinton etc. in HintonG E, Roweis S T.Stochastic neighbor embedding [C] //
Advances in neural information processing systems.2002:833-840. proposing to be known as random close
The visualization Dimension Reduction Analysis method of neighbour embedded (stochastic neighbor embedding, SNE), will be between high dimensional data
Euclidean distance is converted into probability expression-form, and cost functional builds criterion calls subspace and the former input space is having the same
Form of probability.Laurens etc. is in Van der Maaten L, Hinton G.Visualizing data using t-
SNE[J].Journal of Machine Learning Research,2008,9(2579-2605):85. proposing improved
T is distributed SNE, and t-SNE methods use the joint probability expression with symmetry to substitute the conditional probability form in SNE, meanwhile, it is
Solve the problems, such as that data point " crowded " in SNE methods, higher dimensional space are distributed using gaussian probability, lower dimensional space uses degree of freedom
It is distributed for 1 t.This attraction handled in the lower dimensional space for reducing simulation between mapping point.
The algorithm disclose in data classification characteristics, and intuitively expressed by data visualization similar between data
Property.The present invention make full use of t-SNE can disclose in data classification situation the characteristics of instruct the ginseng of NMF preliminary classification classifications
Number setting.By concentrating each model to carry out BOF feature calculations the threedimensional model of input before, therefore can be by the high dimensional data
Collect and carry out visual analyzing as the input of t-SNE, and the results are shown on two-dimensional screen, (a) is shown in Fig. 4, user
The classification situation conduct that by the visual display, the rough classification situation in library can be got information about, and can be observed
Priori, to instruct the parameter setting of NMF preliminary classification classifications, i.e. user NMF can be arranged with observed class label
The initial category value of calculating.Visual advantage can be not only played using this method, but also unsupervised segmentation can be avoided
Blindness.
By being presorted before to Models Sets, therefore input data will have corresponding class label information,
It, should be as far as possible close to similar label data, far from foreign peoples's label data during visualization display.
Models Sets data definition after NMF is presorted isIts
InI-th of sample of c classes is represented, the total classification number of sample is C, and Ni is the sample number of the i-th class, and N=N1+N2+ ...+NC is
Total number of samples.After introducing classification information, former space sample similarity be may be defined as into following form：
Wherein, xiIndicate the feature vector of i-th of model, wherein i, j, k, l, m also in this way, and their dimension by inputting
Model sample number determine, ciIndicate sample xiAffiliated class label information, λ are the variance parameters of corresponding Gaussian function.By
Above-mentioned calculating, above formula maintain the symmetry of probability distribution matrix, and the similarity between homogeneous data and between heterogeneous data
Probability and all be 1.
Similar, the Sample Similarity of subspace is calculated by following formula：
Wherein yiFor sample xiVectorial expression-form in projection subspace.Obtaining similarity pijAnd qijAfterwards, to the greatest extent may be used
The holding of energy with the similarity between class model and reduces the similarity between foreign peoples's sample pattern, and the present invention passes through Kullback-
Leibler divergences can obtain objective cost function：
It farthest reduces the KL divergences of similar sample and increases the KL divergences of foreign peoples's sample.Utilize gradient decline side
Method minimizes the KL divergences of all data points, obtains best simulation point, and the visualization that can complete input data Models Sets is shown,
In Fig. 4 shown in (b).
Step (22) auxiliary area model is shown；
The auxiliary browser window of model is increased, and the display point of auxiliary area model and viewable area carries out one by one
Corresponding display.Visualization is carried out using t-SNE to show, the distribution situation of model is only represented with different color points, is used
Family can not intuitively understand the particular content of the model representated by each point, therefore the visualization for being not appropriate for threedimensional model collection is aobvious
Show.Understanding model hoc scenario in order to facilitate user's profound level and subsequent operation carry out in original display pattern
It improves, that is, increases the auxiliary browser window of model.
User not only can intuitively view the whole classification situation of Models Sets according to original visualization result, also may be used
To be chosen at the corresponding concrete model situation of auxiliary area browsing by region.Furthermore it is also possible to by auxiliary area model with
The display point of viewable area carries out one-to-one correspondence and shows, i.e., chooses corresponding model in auxiliary area, will be in viewable area
Corresponding situation of interior display.
Step (3) changes calculating, Jin Ershi again on the basis of having analysis system by the semi-supervised NMF of design
The now Models Sets dynamic cataloging based on user's driving.This method can support user to carry out Models Sets class on the basis of visual
The operations such as other merging, division.Specifically include the semi-supervised NMF designs of two large divisions and operation：
Step (31) design interacts NMF methods towards threedimensional model collection.
After NMF presorts to Models Sets, matrix centered on obtained W is decomposed, each row of W represent a cluster
Center；H is to refer to matrix, and each row of H represent the generic attribute information of each sample pattern.Merge and divide carrying out Models Sets
It splits in operating process, generates new cluster result.Either merge or division will all provide new reference for system and gather
Class center, therefore can carry out new cluster calculation on this basis and obtain corresponding new classification results.It is calculated when being abstracted into NMF
When upper, the center matrix W obtained by can influencing to presort as the characteristic information of interaction models can be seen as, and be based on gained
The R-matrix of new cluster centre re-start NMF decomposition computations, to obtain completely new cluster result.
Wherein, can interact NMF mathematic(al) representations is：
Formula passes through input parameter V, L, MwIt is best to seek
W, H values.Wherein minW, H >=0It indicates to seek the W under formula value minimizes, H values.
Wherein V is the input matrix for the NMF that step (12) obtains, i.e. threedimensional model collection eigenmatrix；W, H are wanted by NMF
The center matrix sought and reference matrix；L is the R-matrix relative to W；MWFor the diagonal matrix of adjustment parameter, matrix diagonals
Value on line is between 0~1, and α is also adjustment parameter, and wherein range is between 0~1.The meaning of the object function is wishes
The cluster centre that obtained final cluster result can be specified with user is closest.
The operation of step (32) based on semi-supervised NMF.
In the case where the NMF that can be interacted is supported, devises merging and complete relevant mode in real time with the classification continuous iteration of two generic operations
The dynamic of type collection is changed；
Wherein threedimensional model collection categories combination：The present invention supports user by storage optimization, is closed to respective classes
And operate, Fig. 5 is the schematic diagram of a merging process.User can be by browsing come the specified required representative model merged come complete
At corresponding union operation, the model that rectangle frame is identified in (a) in Fig. 5 is just the representative model that need to merge classification, it is noted that
Different models is represented with different shape in schematic diagram of the present invention.It, can be according to class after user specifies the representative model of classification
Attribute information determines its specific column information in Current central matrix W.When merging operating process, it is mainly the desire to root
New cluster centre is obtained according to specified model case, that is, is adjusted and corrected R-matrix L.The present invention in merging process,
The column information not merged in former center matrix W is directly stored in matrix L, then the new cluster for needing pooled model that will be acquired
Center is added in L.The model feature information that the present invention is specified according to user seeks the new column information in R-matrix L
Value, that is, set new cluster centre, for merging two class models, detailed process is as follows：Mould will be specified according to user first
The feature situation of type is sought its closest model neighbour in former class using KNN (k nearest-neighbors) algorithm, is then sought
The center vector (in such as Fig. 5 shown in (b)) of these Model Bs OF features, is added to ginseng as the cluster centre initial value after merging
It examines in matrix L (in such as Fig. 5 shown in (c)).In addition, as also having remaining model after carrying out KNN calculating in specified class, then demand
The center vector (in such as Fig. 5 shown in (b)) for taking remaining model BOF features, as another cluster centre initial value after merging
It is added in R-matrix L (in such as Fig. 5 shown in (c)).After above-mentioned setting, next, just by the object function of setting
Completely new cluster result can be acquired according to the guidance of user.(d) institute in final classification result such as Fig. 5 in Fig. 5 obtained by (a)
Show.
Wherein threedimensional model collection classification divides：The present invention also supports user by storage optimization, is carried out to respective classes
Splitting operation, Fig. 6 are the schematic diagram of a fission process.Division refers to that user creates completely newly on the basis of existing classification results
Classification, i.e. user determine the required classification divided by browsing, and click corresponding representative model, as identified in (a) in Fig. 6
Model, to obtain new classification information.The present invention still completes corresponding operating by reference to matrix L, when division, is equivalent to
Increase new cluster centre on original basis of classification, and then to instruct NMF to carry out completely new calculating.Therefore R-matrix L
To increase the column information for representing new cluster centre on the basis of original center matrix W.The present invention is specified according to user
Model feature information seek the new column information value of R-matrix L, for dividing model I, detailed process is as follows：It is first
First by according to the feature situation of user's designated model, its closest model neighbour in former class is sought using KNN algorithms, then
The center vector (in such as Fig. 6 shown in (b)) for seeking these Model Bs OF features, as the cluster centre initial value addition after division
To ((c) is shown in such as Fig. 6) in R-matrix L.In addition, after carrying out KNN calculating, there is also a need for taking such remaining model BOF special
The center vector (in such as Fig. 6 (b) shown in) of sign, to change column vector information (in Fig. 6 (c) institute of R-matrix L corresponding positions
Show).After above-mentioned setting, next, the object function by setting can acquire completely new cluster according to the guidance of user
As a result.In final classification result such as Fig. 6 in Fig. 6 obtained by (a) shown in (d).
Finally pass through multi-pass operation, obtains final threedimensional model collection classification results as shown in Figure 7.
The threedimensional model sorting technique based on NMF that the present invention provides a kind of, implement the technical solution method and
Approach is all few, the above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill of the art
For personnel, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications
It should be regarded as protection scope of the present invention.All undefined components in this embodiment can be implemented in the prior art.
Claims (1)
1. a kind of threedimensional model sorting technique based on NMF, which is characterized in that include the following steps：
Step 1, input threedimensional model collection, to the threedimensional model collection of input carry out feature calculation obtain NMF initial inputs matrix and
The NMF preliminary classifications of threedimensional model collection are as a result, the threedimensional model that wherein threedimensional model is concentrated is the three-dimensional coordinate for including mesh point
And the triangle grid model of the triangle relation of mesh point；
Step 2, visualization presentation is carried out to threedimensional model collection, wherein visualization refers to after providing visualization interface window for user
Continuous interactive operation；
Step 3, design interacts NMF methods towards threedimensional model collection, according to energy equation, realizes three based on user's driving
Dimension module collection dynamic cataloging, wherein it is respectively to operate division and merging method that can interact NMF methods；
Step 1 includes the following steps：
Step 1-1 builds the BOW bag of words features of threedimensional model collection, obtains NMF initial input matrixes；
Step 1-2 completes the calculating of presorting to threedimensional model collection using NMF, obtains the preliminary classification result of threedimensional model collection；
Step 1-1 includes the following steps：
Step 1-1-1, threedimensional model collection HKS are calculated：It is each to input threedimensional model collection using multi-scale diffusion thermonuclear HKS methods
Mesh point carries out feature calculation, HKS descriptors is obtained, to indicate the local feature information of threedimensional model collection；
Step 1-1-2, vector quantization：The HKS descriptor computations that threedimensional model collection is clustered by k-means obtain corresponding word
Table；
Step 1-1-3 builds the probability distribution that threedimensional model concentrates the correspondence word list of each threedimensional model by statistics, obtains each
The bag of words BOW features of threedimensional model, while obtaining the input matrix V of NMF using calculated BOW features, i.e., three-dimensional mould
Type collection eigenmatrix；
Step 2 includes the following steps：
Step 2-1 is completed the bag of words BOW features of each threedimensional model by t-SNE from higher dimensional space to two dimensional surface
Conversion；
The visualization interface of threedimensional model collection is added in step 2-2, t-SNE conversions projection result is shown to complete defeated
The visualization for entering threedimensional model collection is shown, while carrying out correspondence markings using NMF preliminary classification results to the subpoint of each model,
Visualization display to be completed, subsequent interactive operation is used for, visualization interface window includes auxiliary browser window and viewable area,
The wherein threedimensional model in auxiliary browser window region and the display point of viewable area carries out one-to-one correspondence and shows；
Step 3 includes the following steps：
Step 3-1, design interact NMF energy equations towards threedimensional model collection, and can interact NMF energy equations is：
The minimum value for seeking the expression formula obtains optimal solution, equation
Pass through input parameter V, L, MwTo seek best W, H values, wherein minW,H≥0It indicates to seek the W under formula value minimizes, H values, V
For the input matrix of the obtained NMF of step 1-1-3, i.e. threedimensional model collection eigenmatrix；W, H are respectively in being taken required by NMF
Heart matrix and reference matrix, L are the R-matrix relative to W；MWFor the diagonal matrix of adjustment parameter, matrix MWOn diagonal line
For value between 0~1, a is adjustment parameter, adjusts the weight in equation result convergence direction, range is between 0~1；
Step 3-2 divides two generic operations dynamic change threedimensional model collection by threedimensional model collection categories combination and classification：
Threedimensional model collection categories combination：After the representative threedimensional model of user's specified three-dimensional Models Sets classification, three are represented according to this
Classification belonging to dimension module positions its column information in Current central matrix W, during union operation, according to specified
Threedimensional model obtain new column information, the column information not merged in former center matrix W is directly stored in matrix L, then will
The new cluster centre that the need acquired merge threedimensional model is added in matrix L；
Threedimensional model collection classification divides：User creates completely new classification on the basis of existing classification results, i.e. user passes through browsing
The classification divided needed for determining, and select accordingly to represent the new classification information of obtaining three-dimensional model, when division, in original classification
On the basis of increase new cluster centre, instruct NMF to be calculated, R-matrix L increases generation on the basis of original center matrix W
The column information of the new cluster centre of table.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510695635.2A CN105354593B (en) | 2015-10-22 | 2015-10-22 | A kind of threedimensional model sorting technique based on NMF |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510695635.2A CN105354593B (en) | 2015-10-22 | 2015-10-22 | A kind of threedimensional model sorting technique based on NMF |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105354593A CN105354593A (en) | 2016-02-24 |
| CN105354593B true CN105354593B (en) | 2018-10-30 |
Family
ID=55330560
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510695635.2A Active CN105354593B (en) | 2015-10-22 | 2015-10-22 | A kind of threedimensional model sorting technique based on NMF |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105354593B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106650747B (en) * | 2016-11-16 | 2019-08-02 | 佛山科学技术学院 | A kind of compressed sensing based method for extracting characteristics of three-dimensional model |
| CN106878073B (en) * | 2017-02-14 | 2019-10-08 | 南京邮电大学 | Network multimedia business semisupervised classification method based on t Distribution Mixed Model |
| CN107909099A (en) * | 2017-11-10 | 2018-04-13 | 佛山科学技术学院 | A kind of threedimensional model identification and search method based on thermonuclear |
| CN109034270A (en) * | 2018-08-23 | 2018-12-18 | 西安交通大学 | A kind of visualization feature selection method based on the classification Non-negative Matrix Factorization of failure two |
| CN109492663A (en) * | 2018-09-28 | 2019-03-19 | 西安交通大学 | More characteristic of division selection methods based on Non-negative Matrix Factorization and visualization technique |
| CN109829505A (en) * | 2019-02-15 | 2019-05-31 | 北京工业大学 | A kind of threedimensional model classification method based on various features descriptor |
| CN110704624B (en) * | 2019-09-30 | 2021-08-10 | 武汉大学 | Geographic information service metadata text multi-level multi-label classification method |
| CN113393145B (en) * | 2021-06-25 | 2023-06-30 | 广东利元亨智能装备股份有限公司 | Model similarity obtaining method and device, electronic equipment and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101201845A (en) * | 2007-09-24 | 2008-06-18 | 覃征 | Method for searching three-dimensional model based on virgula point-set delamination helix information |
| CN101320487A (en) * | 2008-07-07 | 2008-12-10 | 中国科学院计算技术研究所 | Scene pretreatment method for fire disaster simulation |
| CN103021029A (en) * | 2013-01-18 | 2013-04-03 | 南京大学 | Automatic labeling method for three-dimensional model component categories |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101221757B1 (en) * | 2011-01-25 | 2013-01-11 | 세종대학교산학협력단 | Method for generating of 3D rule model using decision tree in OLAP and system thereof |
-
2015
- 2015-10-22 CN CN201510695635.2A patent/CN105354593B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101201845A (en) * | 2007-09-24 | 2008-06-18 | 覃征 | Method for searching three-dimensional model based on virgula point-set delamination helix information |
| CN101320487A (en) * | 2008-07-07 | 2008-12-10 | 中国科学院计算技术研究所 | Scene pretreatment method for fire disaster simulation |
| CN103021029A (en) * | 2013-01-18 | 2013-04-03 | 南京大学 | Automatic labeling method for three-dimensional model component categories |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105354593A (en) | 2016-02-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105354593B (en) | A kind of threedimensional model sorting technique based on NMF | |
| Liu et al. | Visualizing high-dimensional data: Advances in the past decade | |
| Li et al. | A comparison of 3D shape retrieval methods based on a large-scale benchmark supporting multimodal queries | |
| CN103207879B (en) | The generation method and apparatus of image index | |
| CN109993748B (en) | Three-dimensional grid object segmentation method based on point cloud processing network | |
| Jänicke et al. | Brushing of attribute clouds for the visualization of multivariate data | |
| CN108875076B (en) | Rapid trademark image retrieval method based on Attention mechanism and convolutional neural network | |
| CN108319957A (en) | A kind of large-scale point cloud semantic segmentation method based on overtrick figure | |
| CN101477529B (en) | Three-dimensional object retrieval method and apparatus | |
| CN106537422A (en) | Systems and methods for capture of relationships within information | |
| CN105243139A (en) | Deep learning based three-dimensional model retrieval method and retrieval device thereof | |
| JPWO2012111236A1 (en) | Image identification apparatus and program | |
| Kalamaras et al. | Multi-objective optimization for multimodal visualization | |
| CN110210431A (en) | A kind of point cloud classifications method based on cloud semantic tagger and optimization | |
| CN109766822A (en) | Gesture identification method neural network based and system | |
| CN110334584B (en) | Gesture recognition method based on regional full convolution network | |
| CN105205135A (en) | 3D (three-dimensional) model retrieving method based on topic model and retrieving device thereof | |
| CN110349159A (en) | 3D shape dividing method and system based on the distribution of weight energy self-adaptation | |
| CN103617609A (en) | A k-means nonlinear manifold clustering and representative point selecting method based on a graph theory | |
| CN110084211A (en) | A kind of action identification method | |
| CN113392244A (en) | Three-dimensional model retrieval method and system based on depth measurement learning | |
| CN106022359A (en) | Fuzzy entropy space clustering analysis method based on orderly information entropy | |
| Wang et al. | Learning to group and label fine-grained shape components | |
| Andreetto et al. | Unsupervised learning of categorical segments in image collections | |
| Wu et al. | Deep texture exemplar extraction based on trimmed T-CNN |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |