CN103021029A - Automatic labeling method for three-dimensional model component categories - Google Patents
Automatic labeling method for three-dimensional model component categories Download PDFInfo
- Publication number
- CN103021029A CN103021029A CN2013100191945A CN201310019194A CN103021029A CN 103021029 A CN103021029 A CN 103021029A CN 2013100191945 A CN2013100191945 A CN 2013100191945A CN 201310019194 A CN201310019194 A CN 201310019194A CN 103021029 A CN103021029 A CN 103021029A
- Authority
- CN
- China
- Prior art keywords
- dimensional model
- dough sheet
- mark
- model
- crf
- 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.)
- Granted
Links
- 238000002372 labelling Methods 0.000 title abstract description 17
- 238000000034 method Methods 0.000 claims abstract description 89
- 238000012549 training Methods 0.000 claims abstract description 73
- 238000007781 pre-processing Methods 0.000 claims abstract description 9
- 238000012216 screening Methods 0.000 claims description 61
- 238000004364 calculation method Methods 0.000 claims description 15
- 230000006870 function Effects 0.000 claims description 15
- 239000012141 concentrate Substances 0.000 claims description 14
- 238000000605 extraction Methods 0.000 claims description 13
- 230000011218 segmentation Effects 0.000 claims description 8
- 238000004422 calculation algorithm Methods 0.000 claims description 5
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 238000012217 deletion Methods 0.000 claims description 3
- 230000037430 deletion Effects 0.000 claims description 3
- 238000005192 partition Methods 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims 2
- 238000010200 validation analysis Methods 0.000 abstract 2
- 238000004458 analytical method Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 230000008901 benefit Effects 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 241001270131 Agaricus moelleri Species 0.000 description 2
- 239000004744 fabric Substances 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000007123 defense Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Abstract
The invention discloses an automatic labeling method for three-dimensional model component categories. The method includes the step of a CRF (conditional random field) labeling model training process, namely, training and learning to obtain a CRF labeling model for segmenting and labeling an unknown three-dimensional model according to a three-dimensional model labeling training set: firstly, dividing the three-dimensional model labeling training set into an instance set and a validation set; then, preprocessing the three-dimensional model labeling training set in the instance set so as to train to obtain unitary items and binary items of the CRF labeling model by classification; and finally, using a preprocessing result of the validation set for parameter search to obtain parameters of the CRF labeling model so that learning of the CRF labeling model is completed. During labeling of the target three-dimensional model, the CRF labeling model obtained from a learning process of the three-dimensional model labeling training set is used for segmenting and labeling, so that category labeling of target three-dimensional model components is achieved.
Description
Technical field
The present invention relates to a kind of disposal route of shape analysis, belong to the computer graphics techniques field, specifically a kind of automatic marking method of three-dimensional model Component Category.
Background technology
Three-dimensional model is divided into the basis that significant component parts is shape understanding and high-level geometric manipulations, further, identifying and obtain the mark problem of describing three-dimensional model component parts classification information is again the key of all multitasks in field such as Geometric Modeling, three-dimensional model animation and texture, for example, in the application that body area network check reason is synthesized, need distinguish in the grid part with " arm " texture or have part of " leg " texture etc.; In addition, some directly do not require to cut apart the application of mark, as, 3D form fit or retrieval also can benefit from component parts and mark classification information.
Although extensive work launches research for automatic image annotation, such as document 1: Bao Hong, Xu Guangmei, Feng Songhe must moral. the automatic image annotation Research progress. and computer science, 2011,38 (7): 35-40., however the work majority of three-dimensional model aspect is only studied for the integral body mark of three-dimensional model, such as document 2: Tian Feng, Shen Xukun, Liu Xianmei, Zhou Kai, Du Ruishan. a kind of three-dimensional model meaning automatic marking method based on weak label, Journal of System Simulation, 2012,24 (9): 1873-1876,1881, and do not relate to the automatic classification mark of three-dimensional model component parts; In addition, the model inseparable with three-dimensional model member mark cut apart the X. such as document 3:Chen, Golovinskiy A., Funkhouser T.A Benchmark for 3D Mesh Segmentation.ACM Transactions on Graphics, 2009,28 (3). described also still is that an opening studies a question, most model dividing methods adopt simple, explainable geometric algorithm, but (for example be subject to general rule, spill, fabric topology, suitable shape primitive) or single feature (for example, the shape diameter, curvature tensor, geodesic distance) divides the input grid, can't be applicable to dissimilar three-dimensional models, more be difficult to realize consistent the cutting apart of similar model classification, thereby be difficult to satisfy the classification mark demand of component parts; Recently, document 4:Golovinskiy A., Funkhouser T.Consistent segmentation of 3D models.Computers and Graphics (Shape Modeling International09) 2009,33 (3): 262-269., document 5: Xu Kai. the 3D shape analysis of semantics-driven and modeling .[D] work such as the .2011. of graduate school of the National University of Defense Technology considers that the three-dimensional model of similar object comprises abundanter semantic information than single model, therefore propose similar Models Sets is analyzed, and then obtain the associating dividing method that a plurality of model consistance are cut apart, yet, the method is not considered the automatic marking problem of unknown three-dimensional model, and can't automatically obtain the classification information of three-dimensional model component parts; Document 6:Kalogerakis E., Hertzmann A., Singh K..Learning 3D mesh segmentation and labeling.ACM Transactions on Graphics, 2010,29 (4) Article No.102. take the lead in proposing a kind of model and cut apart and the learning method that marks, they are by learning the Models Sets of manually cutting apart mark, it is the condition random field optimization problem that model assembly is marked problem representation, thereby realize cutting apart and mark Unknown Model, yet, the learning process of method is consuming time larger, method disclosed by the invention is the further Symmetry Detection by three-dimensional model then, eliminate redundant samples, thus the training time of methods to reduce noises.
Summary of the invention
Goal of the invention: technical matters to be solved by this invention is for the deficiencies in the prior art, and a kind of automatic marking method of three-dimensional model Component Category is provided, and is used for supporting to the auto Segmentation of three-dimensional model and the classification mark of component parts.
In order to solve the problems of the technologies described above, the invention discloses a kind of automatic marking method of three-dimensional model Component Category, may further comprise the steps:
The classification based training process is carried out pre-service to three-dimensional model in the example set, and then classification based training goes out monobasic item and the binary item of CRF marking model.
The parameter search process is searched for dihedral angle parameter lambda, binary probability parameter κ, the length of side parameter μ of only CRF marking model, thereby so that the monobasic item of the CRF marking model that obtains in the combining classification training process and binary item carry out the dough sheet result who marks and the error minimum of verifying the collection mark at the checking collection.
The classification based training part is further comprising the steps of described in the step 1 of the present invention: all three-dimensional models carry out pre-service in the step 111 pair example set, obtain screening dough sheet collection, and the screening dough sheet is concentrated the monobasic proper vector x of each patch grids and the binary feature vector y of adjacent mesh dough sheet.Step 112 adopts the JointBoost sorter that the screening dough sheet is concentrated monobasic proper vector x and the standard mark training study thereof of each patch grids, thereby obtains the monobasic item of CRF marking model.Step 113 adopts the JointBoost sorter screening dough sheet to be concentrated binary feature vector y and the standard mark training study thereof of adjacent mesh dough sheet, calculate again the local geometric features of adjacent mesh, how much dependence G (y) that have mark difference possibility to obtain CRF marking model binary item vacuum metrics adjacent mesh dough sheet, then by the mark compatibility item L (c in the probability calculation CRF marking model binary item that two marks of statistics c is adjacent with c ' in example set, c '), thus finally obtain the binary item of CRF marking model.
Parameter search described in the step 1 of the present invention partly contains and may further comprise the steps: step 121 pair checking concentrates all three-dimensional models to carry out pre-service, obtain screening dough sheet collection, and the screening dough sheet is concentrated the monobasic proper vector x of each patch grids and the binary feature vector y of adjacent mesh dough sheet.Step 122 arranges the value S set (span that generally can set λ, κ, μ is the integer in [0,10]) of λ, κ, μ, and iteration variable i ← 0 is set, and the mark error E is set
m← ∞.Step 123i ← i+1 chooses the i class value λ in the value S set
i, κ
i, μ
i, carry out step 124～125, until all parameters have been chosen in the value S set.Step 124 is according to classification based training step and parameter value λ
i, κ
i, μ
iGained CRF marking model, adopt dough sheet mark process to concentrate the screening dough sheet collection of each three-dimensional model to mark to verifying, and then obtaining the mark that each all patch grids of three-dimensional model is concentrated in checking according to symmetric relation, the computed segmentation weighted error relatively marks error E
mWith error current E
SIf error current is less than mark error: E
S＜E
m, then the value with error current ES is assigned to the mark error E
m, i.e. E
m=E
S, and record optimal parameter { λ
m, κ
m, μ
m} ← { λ
i, κ
i, μ
i, and then return step 123, if error current directly returns step 123 more than or equal to the mark error.
Preprocessing part described in step 1 of the present invention and the step 2 comprises feature extraction and two steps of symmetrical screening: characteristic extraction procedure is analyzed the input three-dimensional model, calculate the monobasic feature of each patch grids, with the binary feature of adjacent mesh dough sheet, described adjacent mesh dough sheet refers to have on the three-dimensional model two patch grids of common edge.Symmetrical screening process is deleted the redundant patch grids in the three-dimensional model according to the symmetry of three-dimensional model, thereby obtains the screening dough sheet collection of three-dimensional model.
Feature extraction part of the present invention is further comprising the steps of: step 11 pair input three-dimensional model carries out standardized operation.At first, the central point with three-dimensional model moves to true origin, the coordinate mean value acquisition of the central point of model by having a few on the computation model; Then, the geodesic distance between any two patch grids on the Calculation of Three Dimensional model; At last, choose the ordering of the size of geodesic distance value between any two patch grids of three-dimensional model in the 30%th geodesic distance value as specification item, with the coordinate of each point on the three-dimensional model divided by this geodesic distance value, thereby finish the standardization of three-dimensional model.Step 12 is extracted the monobasic proper vector of each patch grids of input three-dimensional model, by curvature feature, PCA feature, shape diameter function (shape diameter function, SDF), shape image (volumetric shape image, VSI), average geodesic distance (Average Geodesic Distance, AGD), Shape context (shape contexts, SC), image rotating feature (Spin images) form.Step 13 is extracted the binary feature vector of every group of adjacent mesh dough sheet of each three-dimensional model, is comprised of dihedral angle feature, curvature binary feature, shape diameter difference of function value tag, shape image difference feature.
The symmetrical screening of the present invention part is further comprising the steps of: step 21 adopts voting method to detect the overall situation or the conversion of part reflective symmetry of three-dimensional model, thereby the most significant symmetric pattern in the extraction three-dimensional model, and the symmetric relation between the acquisition dough sheet, described symmetric relation refers in the three-dimensional model whether symmetrical this relation of any two patch grids.Step 22 is put into alternative dough sheet with all dough sheets in the three-dimensional model and is concentrated, and screens the dough sheet collection and be set to empty set.Travel through all dough sheets that alternative dough sheet is concentrated, this dough sheet that traverses put into the screening dough sheet concentrate, and with all with it symmetrical dough sheet concentrate deletion from alternative dough sheet, finally obtain required screening dough sheet collection.
The described dough sheet mark of step 1 of the present invention and step 2 part may further comprise the steps: step 1 makes up a figure, and node of graph is patch grids, has the limit between the adjacent mesh dough sheet, and dough sheet p and dough sheet q are noted as mark f
pWith mark f
qThe time, the geometry in the CRF marking model binary item is relied on the product of a G (y) and the length of side as limit { p, the weights W of q}
P, q(f
p, f
q).Step 2 with CRF marking model monobasic item as data item E
Data(f).Step 3 is with the mark compatibility item L (f in the CRF marking model binary item
p, f
q) as a level and smooth E
Smooth(f) be.Step 4 employing figure cuts optimized algorithm, by calculating the mode of this figure minimal cut, the mark c of each patch grids i of Calculation of Three Dimensional model discrimination dough sheet collection
i
Beneficial effect: the present invention has the following advantages: at first, the present invention can carry out training study to the three-dimensional model mark training set of same kind, obtains the classification marking model of such three-dimensional model; Secondly, the present invention can utilize the classification marking model that goes out from the training set learning that unknown three-dimensional model is carried out significant parts to cut apart, and is not subject to specific rule or feature, thereby is applicable to polytype model segmentation problem; At last, the present invention can utilize the classification marking model that goes out from the training set learning that unknown three-dimensional model is carried out the Component Category automatic marking.
Description of drawings
Below in conjunction with the drawings and specific embodiments the present invention is done further to specify, above-mentioned and/or otherwise advantage of the present invention will become apparent.
Fig. 1 is treatment scheme synoptic diagram of the present invention.
Fig. 2 a and Fig. 2 b are the three-dimensional model mark training set examples of embodiment.
Fig. 3 a is a three-dimensional model synoptic diagram to be marked.
Fig. 3 b is the result schematic diagram that the target three-dimensional model of Fig. 3 a is marked by the CRF marking model of Fig. 2 training set study.
Fig. 4 a is the reflective symmetry floor map that pretreated symmetrical screening step detects a three-dimensional model.
Fig. 4 b is screening dough sheet collection and the monobasic feature binary feature synoptic diagram that preprocessing process obtains.
Fig. 5 a is the result schematic diagram that Fig. 3 a model is marked according to the CRF marking model monobasic item that Fig. 2 a training set classification based training obtains.
Fig. 5 b is the synoptic diagram of CRF marking model binary item on Fig. 3 a model that Fig. 2 a and Fig. 2 b training set classification based training obtain.
Embodiment:
As shown in Figure 1, the automatic marking method of a kind of three-dimensional model Component Category disclosed by the invention specifically may further comprise the steps:
Lower mask body is introduced the main flow process of each step:
1.CRF the training of marking model
The training process of CRF marking model is by learning three-dimensional model mark training set, acquisition can be used for unknown three-dimensional model is cut apart and the CRF marking model that marks, the ratio cut partition of training set according to contained three-dimensional model quantity 4:1 is example set and verifies collection, comprise the classification based training of example set and two steps of parameter search of utilizing checking to collect.
1.1. classification based training
The classification based training process is carried out pre-service to three-dimensional model in the example set, and then classification based training goes out monobasic item and the binary item of CRF marking model.Process is as follows:
Input: three-dimensional model mark example set.
Output: the monobasic item of CRF marking model and binary item.
All three-dimensional models carry out pre-service in the step 1 pair example set, obtain screening dough sheet collection, and the screening dough sheet is concentrated the monobasic proper vector x of each patch grids and the binary feature vector y of adjacent mesh dough sheet.
The monobasic proper vector x of step 2 computation measure patch grids and the conforming CRF marking model of its mark c monobasic item E
1Wherein mark c ∈ C, C is predefined all possible mark set, as, " chair back ", " chair body ", " seat support " or " chair pin ", according to all three-dimensional model screening dough sheet collection and monobasic proper vector x thereof in the example set, and each patch grids marks classification accordingly, adopt document 7:Torralba A., Murphy K.P., Freeman W.T.Sharing Visual Features for Multiclass and Multiview Object Detection.IEEE Transactions On Pattern Analysis and Machine Intelligence, 2007,29 (5): the described JointBoost sorter of 854-869. carries out training study, thereby obtain to have the patch grids of monobasic proper vector x, be labeled as the probability distribution P (c|x) of c, corresponding monobasic Xiang Zewei:
E
1(c;x)=-lo
gP(c|x)
G(y)=-κlog P(c≠c′|y,ξ)-λlog(1-min(ω/π,1)+ε)+μ (1)
Wherein ω is the outer dihedral angle of adjacent mesh dough sheet, ε be have smaller value constant (according to the different accuracy requirement, can be set to 1e-6), thereby avoid calculating log0, λ, κ, μ are dihedral angle parameter, binary probability parameter, the length of side parameter of CRF marking model, can determine its value in the subsequent parameter search procedure.Then, calculate mark compatibility item L (c, c ') to prevent and to be noted as adjacent by adjacent parts, mark compatibility item L (c, c ') obtains by statistics mark c in example set and the adjacent probability calculation of mark c ', the compatibility item value of same mark is 0, if two marks can not be adjacent, then compatible item is set to the maximal value 1 of probable value:
Wherein, a
iBe dough sheet f
iArea.At last, binary Xiang Zewei:
E
2(c,c′;y,λ,κ,μ)=L(c,c′)G(y)(2)
1.2. parameter search
The search of parameter search process is dihedral angle parameter lambda, binary probability parameter κ, the length of side parameter μ of CRF marking model the most accurately, thereby so that the monobasic item of the CRF marking model that obtains in the classification based training process and binary item reach best mark effect at the checking collection.Process is as follows:
Input: three-dimensional model mark checking collection, the monobasic item of CRF marking model and binary item.
Output: the optimal parameter { λ of CRF marking model
m, κ
m, μ
m.
Step 3i ← i+1 chooses the i class value λ in the value S set
i, κ
i, μ
i, carry out step 4～5, until all parameters have been chosen in the value S set.
Wherein mark c
iBe the mark that three-dimensional model gridding dough sheet i obtains by dough sheet mark process, mark
Be the standard mark of three-dimensional model gridding dough sheet i, a
iBe the area of patch grids i,
For verifying that the standard of concentrating is labeled as c
iThe patch grids total area.I (c, c ') is:
Relatively mark error E
mWith error current E
SIf error current is less than mark error: E
S＜E
m, then with error current E
SValue be assigned to the mark error E
m, i.e. E
m=E
S, and record optimal parameter { λ
m, κ
m, μ
m} ← { λ
i, κ
i, μ
i, and then return step 3, if error current directly returns step 3 more than or equal to the mark error.
Finally, the parameter search process obtains the optimal parameter { λ of CRF marking model
m, κ
m, μ
m.
2. the mark of target three-dimensional model
Utilize the learning process gained CRF marking model of three-dimensional model mark training set that the target three-dimensional model is cut apart and mark, thereby obtain the classification mark of target three-dimensional model component parts, this target three-dimensional model consists of for being subordinated to same kind and having same parts with the training lumped model, but not yet cut apart and the three-dimensional model that marks, process is as follows:
Input: target three-dimensional model.
Output: the patch grids mark of target three-dimensional model.
Described preprocessing process is analyzed and is processed the input three-dimensional model, thereby obtains to represent the screening dough sheet collection of all dough sheet features of three-dimensional model, and the feature of each dough sheet wherein.
Input: three-dimensional model.
Output: screening dough sheet collection, and the screening dough sheet is concentrated the monobasic proper vector x of each patch grids and the binary feature vector y of adjacent mesh dough sheet.
Comprise feature extraction and symmetrical two steps of screening.
1. feature extraction
6 pairs of inputs of characteristic extraction procedure list of references three-dimensional model is analyzed, and calculates the monobasic feature of each patch grids, and with the binary feature of adjacent mesh dough sheet, described adjacent mesh dough sheet refers to have on the three-dimensional model two patch grids of common edge.Characteristic extraction procedure is as follows:
(1) curvature feature reference literature 9:Ohtake Y., Belyaev A., Seidel H.-P.Ridge-valley lines on meshes via implicit surface fitting.ACM Transactions on Graphics, 2004,23 (3): the described method of 609-612. is calculated, the match radius adopts respectively on the three-dimensional model between all patch grids 1% of the median of geodesic distance, 2%, 5%, 10%, calculates the gained principal curvatures and uses respectively k
1, k
2Expression so to each match radius value, take principal curvatures as the basis, obtains 9 dimensional features: k
1, | k
1|, k
2, | k
2|, k
1k
2, | k
1k
2|,
k
1-k
2Thereby, the final 36 dimension curvature features that obtain on each dough sheet.
(2) PCA feature calculation process: to each patch grids, calculating its geodesic radius is the covariance matrix of interior all patch grids central points (by the patch grids Area-weighted) of neighborhood of r, and calculates its singular value s
1, s
2, s
3, wherein radius r is respectively 5%, 10%, 20%, 30% and 50% of geodesic distance median between all patch grids.
Wherein, covariance matrix is
Wherein, (x
i, y
i, z
i) be patch grids i center point coordinate,
p
iBe the central point of patch grids i, a
iArea for patch grids i.To each radius value, take singular value as the basis, obtain 12 dimensional features: s
1/ (s
1+ s
2+ s
3), s
2/ (s
1+ s
2+ s
3), s
3/ (s
1+ s
2+ s
3), (s
1+ s
2)/(s
1+ s
2+ s
3), (s
1+ s
3)/(s
1+ s
2+ s
3), (s
2+ s
3)/(s
1+ s
2+ s
3), s
1/ s
2, s
1/ s
3, s
2/ s
3, s
1/ s
2+ s
1/ s
3, s
1/ s
2+ s
2/ s
3, s
1/ s
3+ s
2/ s
3Thereby, the final 60 dimension PCA features that obtain on each dough sheet.
(3) shape diameter Function feature reference literature 10:Shapira L., Shalom S., Shamir A., Cohen-Or D., Zhang H.Contextual part analogies in3D objects.International Journal of Computer Vision, 2010, the described method of 89 (2-3): 309-326. is calculated, wherein the angle parameter of circular cone is set to respectively 15,30,45 degree, to each circular cone, calculate the weighted mean of shape diameter function SDF, median and mean value of square, and then to the logarithm standardization value in these three the value calculating documents 10 of each circular cone, wherein parameter alpha is set to respectively 1,2,4,8, thus final 3 * 3 * 4=36 dimension shape diameter Function feature that obtains on each dough sheet.
(4) shape characteristics of image reference literature 11:Liu R.F., Zhang H., Shamir A., Cohen-Or D.A Part-Aware Surface Metric for Shape Analysis.Computer Graphics Forum (Eurographics 2009), 2009,28 (2): the described method of 397-406. is calculated, wherein all parameters arrange with shape diameter Function feature and calculate, to each circular cone, the same weighted mean of calculating shape image VSI, median and mean value of square, and then calculate logarithm standardization value, thereby final 3 * 3 * 4=36 dimension shape characteristics of image that obtains on each dough sheet.
(5) average geodesic distance feature reference literature 12:Hilaga M., Shinagawa Y., Kohmura T., Kunii T.L.Topology matching for fully automatic similarity estimation of 3d shapes.Proceedings of the 28th annual conference on Computer graphics and interactive techniques (New York, NY, USA, 2001), SIGGRAPH ' 01, ACM, pp.203-212. described method, to each patch grids, calculate the mean value of other patch grids geodesic distances on itself and the three-dimensional model, the 10%th of mean value of square and size ordering, 20%, 90% value, geodesic distance calculates with reference to step 1, thus the final average geodesic distance feature of 11 dimensions that obtains on each dough sheet.
(6) Shape context feature reference literature 13:Belongie S., Malik J., Puzicha J.Shape matching and object recognition using shape contexts.IEEE Transactions On Pattern Analysis and Machine Intelligence, 2002,24 (4): the described method of 509-522., to each patch grids, measure the distribution of other all patch grids (composing power by patch grids self area) in logarithm geodesic distance sector and average angle sector, angle is measured with reference to the patch grids normal direction, adopt 5 geodesic distance sectors and 6 angle sectors, thus the final 30 dimension Shape context features that produce on each dough sheet.
(7) image rotating feature reference literature 14:Johnson A., Hebert M.Using Spin Images for Efficient Object Recognition in Cluttered3D Scenes.IEEE Transactions on Pattern Analysis and Machine Intelligence, 1999,21 (5): the described method of 433-449. is calculated, wherein fix 8 * 8 sector resolution, thus the final 64 dimension image rotating features that produce on each dough sheet.
(1) dihedral angle feature calculation process: the outer dihedral angle ω that calculates adjacent mesh dough sheet i, j
Ij, calculate first feature: dihedral angle numerical characteristics v ω
IjBe value ω
Ij/ π and 1 than decimal:
vω
ij=min(ω
ij/π,1)
On this basis, to every group of adjacent mesh dough sheet i, j, calculate the mean value of all adjacent mesh dough sheet dihedral angle numerical value in its radius r neighborhood, calculate again the index characteristic of these values.Wherein radius r is got 0.5%, 1%, 2%, 4% of geodesic distance median between all patch grids, thereby obtain all the other 4 features, again to these 5 its exponential forms of feature calculation, index value 1,2 ..., 10, thereby the final 50 dimension dihedral angle features that produce on every group of adjacent mesh dough sheet.
(2) curvature binary feature computation process: at first, the curvature value on each summit of Calculation of Three Dimensional model and curvature tensor derivative, reference literature 9, wherein, the match radius is got 0.5% of geodesic distance median between all patch grids, 1%, 2%, 4%, to every group of adjacent mesh dough sheet, calculate curvature derivative on the principal curvatures value of two-end-point on its common edge and the principal direction, and with the mean value of two end points as this adjacent mesh dough sheet binary feature, under each match radius, the curvature derivative respectively has two on principal curvatures value and the principal direction, thereby finally produces 16 dimension curvature binary features on every group of adjacent mesh dough sheet.
(3) shape diameter difference of function value tag computation process: to every group of adjacent mesh dough sheet, calculate the absolute value of two patch grids shape diameter Function feature differences as the binary feature of this adjacent mesh dough sheet, thereby finally produce 36 dimension shape diameter difference of function value tags on every group of adjacent mesh dough sheet.
(4) shape image difference feature calculation process: similarly, to every group of adjacent mesh dough sheet, calculate the absolute value of two patch grids shape characteristics of image differences as the binary feature of this adjacent mesh dough sheet, thereby finally produce 36 dimension shape image difference features on every group of adjacent mesh dough sheet.
2. symmetrical screening
Symmetrical screening process is deleted the redundant patch grids in the three-dimensional model according to the symmetry of three-dimensional model, thereby obtains the screening dough sheet collection of three-dimensional model, and process is as follows:
The target of described dough sheet mark is labeled as c ∈ C for the patch grids that input three-dimensional model screening dough sheet is concentrated marks, and C is predefined all possible mark set, as, " chair back ", " chair body ", " seat support " or " chair pin ", process is as follows:
Input: the CRF marking model comprises monobasic item E
1, binary item E
2(comprise mark compatibility item L (c, c ') and rely on a G (y) how much) and dihedral angle parameter lambda, binary probability parameter κ, length of side parameter μ, three-dimensional model screening dough sheet collection.
Output: the mark c of each patch grids i of this three-dimensional model screening dough sheet collection
i
W
p,q(f
p,f
q)=l
pqG(y
pq,λ,κ,μ)
G (y, λ, κ, μ) is binary item E
2In how much rely on weight l
PqBe the length of adjacent mesh dough sheet common edge, by the standardization of adjacent mesh dough sheet common edge length median, y
PqBinary feature vector for adjacent mesh dough sheet p, q.
Weight a
iBe the area of patch grids p, by the standardization of patch grids area median, x
pMonobasic proper vector for patch grids p.
L (f
p, f
q) be binary item E
2In mark compatibility item.
E(f)=E
data(f)+E
smooth(f)
Optimize this model, and then try to achieve the mark c of each patch grids i of three-dimensional model screening dough sheet collection
i
Embodiment
In the present embodiment, each three-dimensional model in the training set and target three-dimensional model to be marked are carried out preprocessing process.Because present embodiment is the application for three-dimensional model, three-dimensional model different component part can only adopt gray level image to distinguish.At first calculate the monobasic feature of each patch grids by characteristic extraction procedure, binary feature with the adjacent mesh dough sheet, shown in Fig. 4 b, Fig. 4 a is depicted as a three-dimensional model is carried out the reflective symmetry plane that the symmetry screening step of preprocessing process detects for another example, thereby the symmetric relation of utilizing thus reflective symmetry conversion to obtain is deleted the redundant dough sheet in the three-dimensional model, thereby obtain screening dough sheet collection shown in Fig. 4 b, comprising binary feature y and monobasic feature x.
Monobasic feature x and mark according to each dough sheet of example set model discrimination dough sheet collection, adopt the JointBoost sorter to learn out to have the patch grids of monobasic proper vector x, be labeled as the probability distribution P (c|x) of c, thereby obtain the monobasic item of corresponding CRF marking model, mark is shown in Fig. 2 a and Fig. 2 b in the present embodiment: be labeled as " chair back " that marks 1 all dough sheets of part, mark be labeled as " the chair body " of 2 all dough sheets of part, mark be labeled as " seat support " of 3 all dough sheets of part, mark be labeled as " the chair pin " of 4 all dough sheets of part.
Binary feature y and mark according to every group of adjacent mesh dough sheet of example set model discrimination dough sheet collection, adopt the JointBoost sorter learn out the different probability P of adjacent mesh dough sheet mark (c ≠ c ' | y, ξ), thus computation measure adjacent mesh dough sheet exists and to rely on a G (y) how much of mark difference possibility.By the probability calculation mark compatibility item L (c that two marks of statistics c is adjacent with c ' in example set, c '), the mark compatibility item of mark " chair back " and " chair body " is concentrated the ratio calculating of " chair back " dough sheet area summation adjacent with " chair body " and all adjacent surface sheet area summations in the present embodiment by example set screening dough sheet, and mark " chair back " and " chair pin " because can not be adjacent, therefore the compatibility item is set to 1, and the compatibility Xiang Ze of mark " chair back " and " chair back " is set to 0.
The search of parameter search process is dihedral angle parameter lambda, binary probability parameter κ, the length of side parameter μ of CRF marking model the most accurately, thereby so that the monobasic item of the CRF marking model that obtains in the combining classification training process and binary item are minimum with the error of checking collection mark in the result that Fig. 2 b checking collection model carries out the dough sheet mark.
The mark process of target three-dimensional model utilizes above-mentioned CRF marking model that the target three-dimensional model among Fig. 3 a is cut apart and mark, at first obtain screening dough sheet collection and monobasic feature binary feature by preprocessing process, calculate monobasic item and binary item, according to the monobasic item to the mark of model among Fig. 3 a shown in Fig. 5 a, namely, the mark of each dough sheet is made as the mark of monobasic item value minimum, the fringe region mark of circle mark is incorrect, and binary item size is shown in Fig. 5 b, the gray scale smaller part represents that binary item value is less, illustrate that namely the different possibility of adjacent dough sheet mark is larger, also and then by symmetry obtain the annotation results of whole model among Fig. 3 a shown in Fig. 3 b by dough sheet mark process.In the present embodiment, be depicted as the example set of input three-dimensional model mark training set such as four width of cloth figure of Fig. 2 a, Fig. 2 b is the checking collection of input three-dimensional model mark training set, different gray scale charts show and are labeled as different classes of parts, marking 1 part is the chair back, marking 2 parts is the chair body, marking 3 parts is seat support, marking 4 parts is the chair pin, by the automatic marking method of three-dimensional model Component Category of the present invention, can be divided into the parts that represented by different gray scales shown in Fig. 3 b to model among Fig. 3 a, and the final mark that obtains: marking 1 part is the chair back, marking 2 parts is the chair body, and marking 3 parts is seat support, and marking 4 parts is the chair pin.
Claims (7)
1. the automatic marking method of a three-dimensional model Component Category is characterized in that, may further comprise the steps:
Step 1, the training of condition random field CRF marking model: three-dimensional model mark training set is carried out training study, obtain the CRF marking model, described three-dimensional model is the three-dimensional model of patch grids: the three-dimensional model in the three-dimensional model mark training set is to belong to same kind and have the mark three-dimensional model that same parts consists of, wherein markup information is the affiliated classification of three-dimensional model component parts, and is attached on each patch grids of three-dimensional model; It is example set and checking collection according to the ratio cut partition of contained three-dimensional model quantity 4:1 that the learning process of three-dimensional model mark training set marks training set with three-dimensional model, carries out the classification based training of example set and utilizes two steps of parameter search of checking collection:
The classification based training process is carried out pre-service to three-dimensional model in the example set, and then classification based training goes out monobasic item and the binary item of CRF marking model;
The search of parameter search process is dihedral angle parameter lambda, binary probability parameter κ, the length of side parameter μ of CRF marking model the most accurately;
Step 2, the mark of target three-dimensional model: utilize the learning process gained CRF marking model of three-dimensional model mark training set that the target three-dimensional model is cut apart and mark, thereby obtain the classification mark of target three-dimensional model component parts, this target three-dimensional model consists of for being subordinated to same kind and having same parts with the training lumped model, but not yet cuts apart and the three-dimensional model that marks;
Comprise two steps: at first, the target three-dimensional model is carried out pre-service, obtain screening dough sheet collection, and the screening dough sheet is concentrated the monobasic proper vector x of each patch grids and the binary feature vector y of adjacent mesh dough sheet; Then, utilize the CRF marking model that obtains in the step 1, adopt dough sheet mark process that the screening dough sheet collection of target three-dimensional model is marked, and then obtain the mark of all patch grids of target three-dimensional model according to symmetric relation.
2. the automatic marking method of a kind of three-dimensional model Component Category according to claim 1 is characterized in that, preprocessing part described in step 1 and the step 2 comprises feature extraction and two steps of symmetrical screening:
Characteristic extraction procedure is analyzed the input three-dimensional model, calculates the monobasic feature of each patch grids, and with the binary feature of adjacent mesh dough sheet, described adjacent mesh dough sheet refers to have on the three-dimensional model two patch grids of common edge;
Symmetrical screening process is deleted the redundant patch grids in the three-dimensional model according to the symmetry of three-dimensional model, thereby obtains the screening dough sheet collection of three-dimensional model.
3. the automatic marking method of a kind of three-dimensional model Component Category according to claim 2 is characterized in that, feature extraction partly may further comprise the steps:
Step 11, the input three-dimensional model is carried out standardized operation: the central point of three-dimensional model is moved to true origin, and the central point of three-dimensional model obtains by the coordinate mean value of having a few on the computation model; Geodesic distance on the Calculation of Three Dimensional model between any two patch grids; Choose the ordering of the size of geodesic distance value between any two patch grids of three-dimensional model in the 30%th geodesic distance value as specification item, with the coordinate of each point on the three-dimensional model divided by this geodesic distance value, thereby finish the standardization of three-dimensional model;
Step 12 is extracted the monobasic proper vector of inputting each patch grids of three-dimensional model, comprises curvature feature, PCA feature, shape diameter function SDF, shape image VSI, average geodesic distance AGD, Shape context SC, image rotating feature;
Step 13, the binary feature that extracts every group of adjacent mesh dough sheet of each three-dimensional model is vectorial, comprises dihedral angle feature, curvature binary feature, shape diameter difference of function value tag, shape image difference feature.
4. the automatic marking method of a kind of three-dimensional model Component Category according to claim 3 is characterized in that, symmetrical screening may further comprise the steps:
Step 21 adopts voting method to detect the reflective symmetry conversion of three-dimensional model, extracts the most significant symmetric pattern in the three-dimensional model, and obtains the symmetric relation between dough sheet, and described symmetric relation refers to whether any two patch grids are symmetrical in the three-dimensional model;
Step 22 is put into alternative dough sheet with all dough sheets in the three-dimensional model and is concentrated, and screens the dough sheet collection and be set to empty set; Travel through all dough sheets that alternative dough sheet is concentrated, the dough sheet that traverses put into the screening dough sheet concentrate, and with all with it symmetrical dough sheet concentrate deletion from alternative dough sheet, finally obtain required screening dough sheet collection.
5. the automatic marking method of a kind of three-dimensional model Component Category according to claim 4 is characterized in that, the classification based training of example set described in the step 1 may further comprise the steps:
Step 111 is carried out pre-service to all three-dimensional models in the example set, obtains screening dough sheet collection, and the screening dough sheet is concentrated the monobasic proper vector x of each patch grids and the binary feature vector y of adjacent mesh dough sheet;
Step 112 adopts the JointBoost sorter to concentrate monobasic proper vector x and the mark thereof of each patch grids to carry out training study to the screening dough sheet, thereby obtains the monobasic item of CRF marking model;
Step 113, adopt the JointBoost sorter to concentrate binary feature vector y and the mark thereof of adjacent mesh dough sheet to carry out training study to the screening dough sheet, calculate the local geometric features of adjacent mesh, obtain CRF marking model binary item vacuum metrics adjacent mesh dough sheet and exist the geometry of mark difference possibility to rely on a G (y); At any two the mark c probability adjacent with c ' of example centralized calculation, calculate the mark compatibility item L (c, c ') in the CRF marking model binary item, finally obtain the binary item of CRF marking model.
6. the automatic marking method of a kind of three-dimensional model Component Category according to claim 5 is characterized in that, the parameter search of the collection of checking described in the step 1 may further comprise the steps:
Step 121 concentrates all three-dimensional models to carry out pre-service to verifying, obtains screening dough sheet collection, and the screening dough sheet is concentrated the monobasic proper vector x of each patch grids and the binary feature vector y of adjacent mesh dough sheet;
Step 122 arranges the value S set of dihedral angle parameter lambda, binary probability parameter κ, length of side parameter μ, and the span of λ, κ, μ is the integer in [0,10], and it is 0 that iteration variable i is set, and the mark error E is set
mBe infinity;
Step 123, iteration variable i increases by 1, chooses the i group parameter value λ in the value S set
i, κ
i, μ
i, carry out step 124～125, until all parameters have been chosen in the value S set;
Step 124 is according to classification based training and the parameter value λ of example set
i, κ
i, μ
iGained CRF marking model adopts dough sheet mark process to concentrate the screening dough sheet collection of each three-dimensional model to mark to verifying, and then obtains the mark that each all patch grids of three-dimensional model is concentrated in checking, computed segmentation weighted error according to symmetric relation; Relatively mark error E
mWith error current E
SIf error current is less than mark error: E
S＜E
m, then with error current E
SValue be assigned to the mark error E
m, i.e. E
m=E
S, and the most accurate parameter { λ of record
m, κ
m, μ
m, and then return step 123, if error current directly returns step 123 more than or equal to the mark error.
7. the automatic marking method of a kind of three-dimensional model Component Category according to claim 6 is characterized in that, the described dough sheet mark of step 1 and step 2 part may further comprise the steps:
Step 1 makes up a figure, and node of graph is patch grids, has the limit between the adjacent mesh dough sheet, and dough sheet p and dough sheet q are noted as mark f
pWith mark f
qThe time, rely on the product of a G (y) and the length of side as common edge { p, the weights W of q} of dough sheet p and dough sheet q with how much in the CRF marking model binary item
P, q(f
p, f
q);
Step 2, with CRF marking model monobasic item as data item E
Data(f);
Step 3 is with the mark compatibility item L (f in the CRF marking model binary item
p, f
q) as a level and smooth E
Smooth(f);
Step 4, employing figure cuts optimized algorithm, by calculating the mode of this figure minimal cut, the mark c of each patch grids i of Calculation of Three Dimensional model discrimination dough sheet collection
i
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310019194.5A CN103021029B (en) | 2013-01-18 | 2013-01-18 | Automatic labeling method for three-dimensional model component categories |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310019194.5A CN103021029B (en) | 2013-01-18 | 2013-01-18 | Automatic labeling method for three-dimensional model component categories |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103021029A true CN103021029A (en) | 2013-04-03 |
| CN103021029B CN103021029B (en) | 2015-03-25 |
Family
ID=47969594
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310019194.5A Expired - Fee Related CN103021029B (en) | 2013-01-18 | 2013-01-18 | Automatic labeling method for three-dimensional model component categories |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103021029B (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103473813A (en) * | 2013-09-18 | 2013-12-25 | 南京大学 | Automatic extracting method for three-dimensional model members |
| CN103838933A (en) * | 2014-03-18 | 2014-06-04 | 湖南大学 | Three-dimensional model integrating and labeling method based on digital technology |
| CN103971415A (en) * | 2014-05-23 | 2014-08-06 | 南京大学 | Online marking method for three-dimensional model component |
| CN105354593A (en) * | 2015-10-22 | 2016-02-24 | 南京大学 | NMF (Non-negative Matrix Factorization)-based three-dimensional model classification method |
| CN106157375A (en) * | 2016-07-06 | 2016-11-23 | 南京大学 | A kind of threedimensional model component categories automatic marking method |
| CN107657279A (en) * | 2017-09-26 | 2018-02-02 | 中国科学院大学 | A kind of remote sensing target detection method based on a small amount of sample |
| CN108021683A (en) * | 2017-12-11 | 2018-05-11 | 首都航天机械公司 | A kind of scale model retrieval implementation method based on three-dimensional labeling |
| CN108121862A (en) * | 2017-12-13 | 2018-06-05 | 武汉益模科技股份有限公司 | A kind of engineering drawing automatic marking method based on three-dimensional geometry feature |
| CN108229772A (en) * | 2016-12-14 | 2018-06-29 | 北京国双科技有限公司 | Mark processing method and processing device |
| CN109815979A (en) * | 2018-12-18 | 2019-05-28 | 通号通信信息集团有限公司 | A kind of weak label semantic segmentation nominal data generation method and system |
| CN109976614A (en) * | 2019-03-28 | 2019-07-05 | 广州视源电子科技股份有限公司 | A kind of mask method of solid figure, device, equipment and medium |
| CN110533781A (en) * | 2019-08-28 | 2019-12-03 | 南京信息职业技术学院 | A kind of multi-class threedimensional model component automatic marking method |
| CN110689623A (en) * | 2019-08-20 | 2020-01-14 | 重庆特斯联智慧科技股份有限公司 | Tourist guide system and method based on augmented reality display |
| CN111462326A (en) * | 2020-03-31 | 2020-07-28 | 武汉大学 | Low-cost 360-degree panoramic video camera urban pipeline three-dimensional reconstruction method and system |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102163340A (en) * | 2011-04-18 | 2011-08-24 | 宁波万里电子科技有限公司 | Method for labeling three-dimensional (3D) dynamic geometric figure data information in computer system |
-
2013
- 2013-01-18 CN CN201310019194.5A patent/CN103021029B/en not_active Expired - Fee Related
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102163340A (en) * | 2011-04-18 | 2011-08-24 | 宁波万里电子科技有限公司 | Method for labeling three-dimensional (3D) dynamic geometric figure data information in computer system |
Non-Patent Citations (3)
| Title |
|---|
| BOYKOV, Y.,VEKSLER, O.: "Fast Approximate Energy Minimization via Graph Cuts", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》, 7 August 2002 (2002-08-07) * |
| EVANGELOS KALOGERAKIS等: "Learning 3D Mesh Segmentation and Labeling", 《ACM TRANSACTIONS ON GRAPHICS》, 31 July 2010 (2010-07-31) * |
| NILOY J. MITRA等: "Partial and Approximate Symmetry Detection for 3D Geometry", 《ACM TRANSACTIONS ON GRAPHICS》, 31 July 2006 (2006-07-31) * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103473813A (en) * | 2013-09-18 | 2013-12-25 | 南京大学 | Automatic extracting method for three-dimensional model members |
| CN103473813B (en) * | 2013-09-18 | 2016-06-15 | 南京大学 | A kind of extraction method of threedimensional model component |
| CN103838933A (en) * | 2014-03-18 | 2014-06-04 | 湖南大学 | Three-dimensional model integrating and labeling method based on digital technology |
| CN103838933B (en) * | 2014-03-18 | 2017-02-15 | 湖南大学 | Three-dimensional model integrating and labeling method based on digital technology |
| CN103971415A (en) * | 2014-05-23 | 2014-08-06 | 南京大学 | Online marking method for three-dimensional model component |
| CN103971415B (en) * | 2014-05-23 | 2016-06-15 | 南京大学 | The online mask method of a kind of three-dimensional model component |
| CN105354593A (en) * | 2015-10-22 | 2016-02-24 | 南京大学 | NMF (Non-negative Matrix Factorization)-based three-dimensional model classification method |
| CN105354593B (en) * | 2015-10-22 | 2018-10-30 | 南京大学 | A kind of threedimensional model sorting technique based on NMF |
| CN106157375A (en) * | 2016-07-06 | 2016-11-23 | 南京大学 | A kind of threedimensional model component categories automatic marking method |
| CN106157375B (en) * | 2016-07-06 | 2018-11-30 | 南京大学 | A kind of threedimensional model component categories automatic marking method |
| CN108229772A (en) * | 2016-12-14 | 2018-06-29 | 北京国双科技有限公司 | Mark processing method and processing device |
| CN107657279A (en) * | 2017-09-26 | 2018-02-02 | 中国科学院大学 | A kind of remote sensing target detection method based on a small amount of sample |
| CN107657279B (en) * | 2017-09-26 | 2020-10-09 | 中国科学院大学 | Remote sensing target detection method based on small amount of samples |
| CN108021683A (en) * | 2017-12-11 | 2018-05-11 | 首都航天机械公司 | A kind of scale model retrieval implementation method based on three-dimensional labeling |
| CN108021683B (en) * | 2017-12-11 | 2021-06-22 | 首都航天机械公司 | Three-dimensional labeling-based similar model retrieval implementation method |
| CN108121862A (en) * | 2017-12-13 | 2018-06-05 | 武汉益模科技股份有限公司 | A kind of engineering drawing automatic marking method based on three-dimensional geometry feature |
| CN108121862B (en) * | 2017-12-13 | 2021-04-27 | 武汉益模科技股份有限公司 | Automatic engineering drawing labeling method based on three-dimensional geometric features |
| CN109815979A (en) * | 2018-12-18 | 2019-05-28 | 通号通信信息集团有限公司 | A kind of weak label semantic segmentation nominal data generation method and system |
| CN109976614A (en) * | 2019-03-28 | 2019-07-05 | 广州视源电子科技股份有限公司 | A kind of mask method of solid figure, device, equipment and medium |
| CN109976614B (en) * | 2019-03-28 | 2021-04-06 | 广州视源电子科技股份有限公司 | Method, device, equipment and medium for marking three-dimensional graph |
| CN110689623A (en) * | 2019-08-20 | 2020-01-14 | 重庆特斯联智慧科技股份有限公司 | Tourist guide system and method based on augmented reality display |
| CN110533781A (en) * | 2019-08-28 | 2019-12-03 | 南京信息职业技术学院 | A kind of multi-class threedimensional model component automatic marking method |
| CN111462326A (en) * | 2020-03-31 | 2020-07-28 | 武汉大学 | Low-cost 360-degree panoramic video camera urban pipeline three-dimensional reconstruction method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103021029B (en) | 2015-03-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103021029A (en) | Automatic labeling method for three-dimensional model component categories | |
| CN104091321B (en) | It is applicable to the extracting method of the multi-level point set feature of ground laser radar point cloud classifications | |
| US10013507B2 (en) | Learning synthetic models for roof style classification using point clouds | |
| Li et al. | PlantNet: A dual-function point cloud segmentation network for multiple plant species | |
| Appleton et al. | Globally optimal geodesic active contours | |
| CN102004922B (en) | High-resolution remote sensing image plane extraction method based on skeleton characteristic | |
| Noma et al. | Interactive image segmentation by matching attributed relational graphs | |
| CN106157375B (en) | A kind of threedimensional model component categories automatic marking method | |
| Bhowmick et al. | Divide and conquer: Efficient large-scale structure from motion using graph partitioning | |
| Miao et al. | A method for accurate road centerline extraction from a classified image | |
| CN109002834A (en) | Fine granularity image classification method based on multi-modal characterization | |
| CN103745191B (en) | Landform analysis based method for automatically identifying tablelands, ridges and loess hills in loess region | |
| CN103077555A (en) | Automatic marking method composed of three-dimensional models | |
| CN110222642A (en) | A kind of planar architectural component point cloud contour extraction method based on global figure cluster | |
| Poux et al. | Self-learning ontology for instance segmentation of 3d indoor point cloud | |
| Wang et al. | Feature extraction of point clouds based on region clustering segmentation | |
| CN103955709B (en) | Weighted synthetic kernel and triple markov field (TMF) based polarimetric synthetic aperture radar (SAR) image classification method | |
| CN103971415A (en) | Online marking method for three-dimensional model component | |
| Tardy et al. | Uncertainty measurements for the reliable classification of mammograms | |
| Khan et al. | A modified adaptive differential evolution algorithm for color image segmentation | |
| CN103473813A (en) | Automatic extracting method for three-dimensional model members | |
| CN113591016B (en) | Landslide labeling contour generation method based on multi-user cooperation | |
| CN104463825B (en) | Device and method for the test object in three-dimensional volumetric images | |
| CN106980878B (en) | Method and device for determining geometric style of three-dimensional model | |
| CN110334704B (en) | Three-dimensional model interest point extraction method and system based on layered learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20150325 Termination date: 20180118 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |