CN107092859A - A kind of depth characteristic extracting method of threedimensional model - Google Patents
A kind of depth characteristic extracting method of threedimensional model Download PDFInfo
- Publication number
- CN107092859A CN107092859A CN201710148547.XA CN201710148547A CN107092859A CN 107092859 A CN107092859 A CN 107092859A CN 201710148547 A CN201710148547 A CN 201710148547A CN 107092859 A CN107092859 A CN 107092859A
- Authority
- CN
- China
- Prior art keywords
- mrow
- layer
- msub
- neural networks
- convolutional neural
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/64—Three-dimensional objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of depth characteristic extracting method of threedimensional model, first, extracts the pole view of threedimensional model, is used as the training input data of depth convolutional neural networks；Secondly, depth convolutional neural networks are built, and pole view is trained；Again, pole view is input into depth convolutional neural networks to be trained, until the convergence of depth convolutional neural networks, realizes the determination of the internal weights of depth convolutional neural networks after the completion of training；Finally, input need to extract the pole view of the threedimensional model of feature in the depth convolutional neural networks trained, calculate the characteristic vector of full linking layer in depth convolutional neural networks, then as the depth characteristic for the threedimensional model that need to extract feature.The present invention builds deep layer convolutional neural networks and reduces residual error by iterated revision weights so that network convergence.After training is finished, the full linking layer for extracting convolutional neural networks is used as the depth characteristic of threedimensional model pole view.
Description
Technical field
The present invention relates to threedimensional model processing technology field, more specifically to a kind of depth characteristic of threedimensional model
Extracting method.
Background technology
With threedimensional model treatment technology and computer hardware, the fast development of software, and multimedia technology and interconnection
The popularization of network technology, substantial amounts of threedimensional model is applied in every field, and people also increasingly increase to the application demand of threedimensional model
Greatly.Threedimensional model plays the part of important in multiple fields such as ecommerce, architectural design, industrial design, advertisement video display and 3d gamings
Role.The threedimensional model of large-scale dataset needs design reuse and model index in the various aspects that social production is lived,
Therefore, how to be concentrated from existing all types of three-dimensional modeling data and quickly and accurately retrieve target three-dimensional, become
Key issue urgently to be resolved hurrily.
In recent years, the threedimensional model analysis based on deep learning turns into study hotspot, and it combines computer vision, artificial
The research contents such as intelligence, intelligence computation, can solve the visual task of threedimensional model includes threedimensional model feature extraction, classification, knows
Not, the problem such as detection and prediction.The hidden feature of threedimensional model can automatically be learnt using depth learning technology, and can be big
It is trained on scale data collection, strengthens the generalization ability of learning model.
The feature extracting method of deep learning is currently based on, there are the following problems：The spy that framework such as deep learning is extracted
Levy and can not express three-dimensional model information completely, high and network the over-fitting problem of computation complexity that network layer bathozone is come, with
And e-learning time length and memory storage space it is big etc..Used and to threedimensional model feature with the technology maturation of deep learning
Capability list reaches strong demand, and it will be threedimensional model classification, retrieval, detection and identification problem band to go to extract feature using deep learning
Carry out new breakthrough.
The content of the invention
It is an object of the invention to overcome shortcoming and deficiency of the prior art, there is provided a kind of depth characteristic of threedimensional model
Extracting method, this method builds deep layer convolutional neural networks and the pole view of threedimensional model is trained, and is weighed by iterated revision
Value reduces residual error so that network convergence.After training is finished, the full linking layer for extracting convolutional neural networks is used as threedimensional model pole
The depth characteristic of view so that depth characteristic can be used for the visual tasks such as the classification, retrieval and identification of threedimensional model.This method structure
The convolutional neural networks level for building deep layer enriches, and can accelerate network training speed and improve the degree of accuracy of depth network fitting.
In order to achieve the above object, the technical scheme is that：A kind of depth of threedimensional model is special
Levy extracting method, it is characterised in that：
First, the pole view of threedimensional model is extracted, the training input data of depth convolutional neural networks is used as；
Secondly, depth convolutional neural networks are built, and pole view is trained；Wherein, depth convolutional neural networks bag
Include pole view as training input data input layer, learnt for the feature to pole view and obtain two dimensional character figure
Convolutional layer, for the two dimensional character figure of diverse location is carried out aggregate statistics and reduce characteristic dimension pond layer, for pair
Two dimensional character figure carries out arrangement link and forms the full linking layer of one-dimensional vector and for exporting the output that classification predicts the outcome
Layer；
Again, pole view is input into depth convolutional neural networks to be trained, until the convergence of depth convolutional neural networks,
Realize the determination of the internal weights of depth convolutional neural networks after the completion of training；
Finally, input need to extract the pole view of the threedimensional model of feature in the depth convolutional neural networks trained,
Calculate the characteristic vector of full linking layer in depth convolutional neural networks, then it is special as the depth for the threedimensional model that need to extract feature
Levy.
In such scheme, the depth characteristic extracting method of threedimensional model of the invention is by depth convolutional neural networks
The pole view of threedimensional model is trained, residual error is reduced by iterated revision weights so that network convergence.Treat that training is finished
Afterwards, the full linking layer for extracting convolutional neural networks is used as the depth characteristic of threedimensional model pole view.Wherein, pole view is to three-dimensional mould
The spatial aggregation structure global table of type is reached, and the instruction of depth convolutional neural networks can be simplified and mitigate using the pole view of threedimensional model
Practice amount of calculation.Depth convolutional neural networks level constructed by the present invention enriches, and can accelerate network training speed and improve depth
The degree of accuracy of network fitting.Wherein, pole view refers to outwards launch one group of sampling ray, ray and mould from the barycenter of threedimensional model
The two-dimentional sample graph that the distance of the intersection point of type to barycenter is arranged in.
Specifically, this method comprises the following steps：
Step s101：The pole view of threedimensional model is extracted, as the training input data of depth convolutional neural networks, wherein
Training input data is set to x(i)∈ χ, χ are the pole view for having N number of threedimensional model；The corresponding class label of i-th of model is
y(i)∈ { 1,2 ..., K }, K are the categorical measure of threedimensional model；
Step s102：Build depth convolutional neural networks；In depth convolutional neural networks, depth convolutional neural networks bag
Include：Pole view x(i)It is used as the input layer Ι, 4 convolutional layer C (t), t=1,2 of training input data, 3,4,1 pond layer P, two
Individual full linking layer FC (1) and FC (2), and output layer Ο；Wherein, 4 convolutional layer C (t), t=1,2, each convolutional layer in 3,4
With each full linking layer in two full linking layer FC (1) and FC (2) using correct linear activation primitive f (a)=max (0, a)
Substitute sigmoid function f (a)=1/ (1+e in depth convolutional neural networks-a)；
Step s103：The parameter of depth convolutional neural networks is set, i.e., it is each layer weights of depth convolutional neural networks are initial
Change：
Input layer Ι：Input data is the pole view x that a size is (32 × 32)(i)；
Convolutional layer C：Feature of 4 convolutional layers successively to pole view learns, the two dimensional character that each convolutional layer is obtained
The quantity of figure is Ft=(6,8,10,12), the two dimensional character figure of each convolutional layer is obtained according to below equation：
WhereinC (t) q-th of two dimensional character figure is represented, M represents t-1 layers of feature set of graphs, when t-1 is 0, i.e. table
Show the pole view that two dimensional character figure is input,Represent t-1 layers of p-th of characteristic pattern to q-th of two dimension of t layers of convolutional layer
The convolution kernel of characteristic pattern, initial convolution kernel is used as by the matrix of the generating random number (5 × 5) of [- 1,1].Bias is biasing, initially
Change value is 0.(*) represents the computing of convolution, and f () is the linear activation primitive of amendment；It is available by above-mentioned formula：
Convolutional layer C (1) calculates the two dimensional character figure that 6 sizes are (28 × 28)；
Convolutional layer C (2) calculates the two dimensional character figure that 8 sizes are (24 × 24)；
Convolutional layer C (3) calculates the two dimensional character figure that 10 sizes are (20 × 20)；
Convolutional layer C (4) calculates the two dimensional character figure that 12 sizes are (16 × 16)；
Pond layer P：The two dimensional character figure calculated by following formula to convolutional layer C (4) carries out maximum pondization and handled, and obtains
To the eigenmatrix that 12 sizes are (8 × 8)：
Wherein, PMp(u0,v0) the corresponding coordinate of eigenmatrix that the maximum pondization of two dimensional character figure is handled is represented,It is convolution
The two dimensional character figure that layer C (4) is calculated, max () is the maximum of calculating matrix element；
Full linking layer FC (1)：By PMpColumn vector (64 × 1) is arranged in, one is obtained to the full link of each matrix column vector
Dimensional vector L0(768 × 1), i.e. one-dimensional vector L0(768 × 1) as full linking layer FC (1) input vector；
Full linking layer FC (2)：Setting full linking layer FC (1) has 512 neurons, and full linking layer FC (2) has 128 nerves
Member；And the output vector L of full linking layer FC (1) is calculated by following full linking layer propagation formulas1It is used as full linking layer FC (2)
Input vector, and calculate the output vector L of full linking layer FC (2)2It is used as output layer O input vector：
WhereinForWithNetwork weight, is made by the matrix of the generating random number (512 × 768) of [- 1,1]
For initial weight W1, generation (128 × 512) matrix be used as initial weight W2；ForBiasing, initial value is 0；f
() is the linear activation primitive of amendment；Take 1 and 2；
Output layer O：Output layer is set to have K neuron, calculation formula is as follows：
Y '=f (W3L2+b2)；
Wherein, y '(i)For the final output of depth convolutional neural networks, i is i-th of model；W3Random number by [- 1,1]
The matrix of generation (K × 128) is used as initial weight；L2For the output vector of full linking layer FC (2)；
Step s104：The data set χ of pole view is trained, it is 1 to set learning rate η, declines calculation at random using gradient
Method, the result y ' of prediction is exported according to depth convolutional neural networks(i)With real class label y(i)∈ { 1,2 ..., K } mistake
Difference, carries out backpropagation, and 20 algorithms of iteration can restrain, and realizes the internal weights of depth convolutional neural networks after the completion of training
It is determined that；
Step s105：The pole that the threedimensional model that need to extract feature is inputted in the depth convolutional neural networks trained is regarded
Scheme x(i), and calculate the characteristic vector L of the 2nd full linking layer FC (2) outputs2, the threedimensional model pole that feature is extracted needed for being is regarded
The depth characteristic of figure.
The pole view for extracting threedimensional model refers to：
First, three-dimensional point cloud model is pre-processed, calculates the barycenter and yardstick of three-dimensional point cloud model, and by three-dimensional
Point cloud model is moved in rectangular coordinate system and zoomed in and out, and realizes that three-dimensional point cloud model fastens normalization in rectangular co-ordinate；
Secondly, spherical coordinates is transformed into by the three-dimensional point cloud model of scaling by being fastened in rectangular co-ordinate, and obtain three-dimensional point
The direction of each point of cloud model and distance property；
Again, the spherical coordinates of point set is mapped on the location of pixels of pole view, calculates each pixel sampling distance set
Ultimate range, is used as the ray sampled value of Direction interval；
Finally, the ultimate range of each pixel sampling distance set is arranged in two-dimentional sample graph, is that extracted pole is regarded
Figure.
The pole view for extracting threedimensional model specifically includes following steps：
Step S201：Three-dimensional point cloud model is inputted, the yardstick of the three-dimensional point cloud model is P={ pi(xi,yi,zi) | i=1,
2,...,N}；
Step S202：Barycenter g (the g of three-dimensional point cloud model are calculated according to the following equationx, gy, gz), pass through obtained barycenter g
(gx, gy, gz) fasten three-dimensional point cloud model translation transformation to rectangular co-ordinate, then chi of the three-dimensional point cloud model in rectangular coordinate system
Spend for pi'=pi- g, i=1,2 ..., N；Three-dimensional point cloud model p after translation transformationi' barycenter be located at rectangular coordinate system original
Point；
Step s203：The zoom factor s of three-dimensional point cloud model is calculated, three-dimensional point cloud model is zoomed into Unit Scale isOn, wherein, zoom factor s is
Step s204：Spherical coordinates Q is transformed into by the three-dimensional point cloud model of scaling by being fastened in rectangular co-ordinate, it is now three-dimensional
Yardstick of the point cloud model on spherical coordinates beConversion formula is as follows：
Wherein θ ∈ [0, π], it is 0 that the elevation angle is born on semiaxis in Z axis；
Step s205：Spherical coordinates Q is mapped on the location of pixels (u, v) of pole view, mapping relations areThen three-dimensional point cloud model is calculated according to the following equation on the location of pixels (u, v) of pole view；Wherein,
There is spherical coordinates point, multiple spherical coordinates point on one location of pixels (u, v) or in the absence of there is spherical coordinates point；
Wherein nuAnd nvRespectively pole view is wide and long；
Step s206：Each the sampled distance collection of pixel (u, v) isGo out each picture according to the following formula
The maximum of plain sampled distance collection, to obtain pixel sampling value in the view of pole, and is arranged in two-dimentional collection figure and made as ultimate range
For pole view I；
Above-mentioned depth characteristic extracting method directly carries out feature extraction to the pole view of threedimensional model, is regarded for two-dimentional pole
The information of each position can be perceived with convolution operation by desiring to make money or profit.The characteristic pattern obtained by multilayer process of convolution, in full linking layer
It is middle to extract the strong depth characteristic of discrimination property.
And when extracting pole view, preprocessing process needs to carry out threedimensional model Pan and Zoom conversion, it is ensured that three-dimensional
Model is normalized and standardization in standard scale.Threedimensional model point cloud is converted into spherical coordinate system, beneficial to by spherical coordinates
Point is mapped to the respective pixel position of two-dimentional pole view, by mapping, counts the maximum of the distance set of point on the location of pixels
Value, forms two-dimentional sample graph by maximum sampled value, is the new pole view of threedimensional model.
Compared with prior art, the invention has the advantages that and beneficial effect：The depth characteristic of threedimensional model of the present invention
Extracting method builds deep layer convolutional neural networks and the pole view of threedimensional model is trained, and reduces residual by iterated revision weights
Difference so that network convergence.After training is finished, the full linking layer for extracting convolutional neural networks is used as the depth of threedimensional model pole view
Spend feature so that depth characteristic can be used for the visual tasks such as the classification, retrieval and identification of threedimensional model.This method builds deep layer
Convolutional neural networks level enriches, and can accelerate network training speed and improve the degree of accuracy of depth network fitting.
Brief description of the drawings
Fig. 1 is the flow chart of the depth characteristic extracting method of threedimensional model of the present invention；
Fig. 2 is the schematic diagram of the depth characteristic extracting method mid-deep strata convolutional neural networks of threedimensional model of the present invention；
Fig. 3 is the flow chart of the pole view of extraction threedimensional model in the inventive method；
Fig. 4 is to extract the schematic diagram for obtaining pole view in the inventive method by threedimensional model；
Embodiment
The present invention is described in further detail with embodiment below in conjunction with the accompanying drawings.
Embodiment
As shown in Figures 1 to 4, the depth characteristic extracting method of threedimensional model of the present invention is such：
First, the pole view of threedimensional model is extracted, the training input data of depth convolutional neural networks is used as；
Secondly, depth convolutional neural networks are built, and pole view is trained；Wherein, depth convolutional neural networks bag
Include pole view as training input data input layer, learnt for the feature to pole view and obtain two dimensional character figure
Convolutional layer, for the two dimensional character figure of diverse location is carried out aggregate statistics and reduce characteristic dimension pond layer, for pair
Two dimensional character figure carries out arrangement link and forms the full linking layer of one-dimensional vector and for exporting the output that classification predicts the outcome
Layer；
Again, pole view is input into depth convolutional neural networks to be trained, until the convergence of depth convolutional neural networks,
Realize the determination of the internal weights of depth convolutional neural networks after the completion of training；
Finally, input need to extract the pole view of the threedimensional model of feature in the depth convolutional neural networks trained,
Calculate the characteristic vector of full linking layer in depth convolutional neural networks, then it is special as the depth for the threedimensional model that need to extract feature
Levy.
In such scheme, the depth characteristic extracting method of threedimensional model of the invention is by depth convolutional neural networks
The pole view of threedimensional model is trained, residual error is reduced by iterated revision weights so that network convergence.Treat that training is finished
Afterwards, the full linking layer for extracting convolutional neural networks is used as the depth characteristic of threedimensional model pole view.Wherein, pole view is to three-dimensional mould
The spatial aggregation structure global table of type is reached, and the instruction of depth convolutional neural networks can be simplified and mitigate using the pole view of threedimensional model
Practice amount of calculation.Depth convolutional neural networks level constructed by the present invention enriches, and can accelerate network training speed and improve depth
The degree of accuracy of network fitting.Wherein, pole view refers to outwards launch one group of sampling ray, ray and mould from the barycenter of threedimensional model
The two-dimentional sample graph that the distance of the intersection point of type to barycenter is arranged in.
Specifically, this method comprises the following steps：
Step s101：The pole view of threedimensional model is extracted, as the training input data of depth convolutional neural networks, wherein
Training input data is set to x(i)∈ χ, χ are the pole view for having N number of threedimensional model；The corresponding class label of i-th of model is
y(i)∈ { 1,2 ..., K }, K are the categorical measure of threedimensional model；
Step s102：Build depth convolutional neural networks；In depth convolutional neural networks, depth convolutional neural networks bag
Include：Pole view x(i)It is used as the input layer Ι, 4 convolutional layer C (t), t=1,2 of training input data, 3,4,1 pond layer P, two
Individual full linking layer FC (1) and FC (2), and output layer Ο；Wherein, 4 convolutional layer C (t), t=1,2, each convolutional layer in 3,4
With each full linking layer in two full linking layer FC (1) and FC (2) using correct linear activation primitive f (a)=max (0, a)
Substitute sigmoid function f (a)=1/ (1+e in depth convolutional neural networks-a)；
Step s103：The parameter of depth convolutional neural networks is set, i.e., it is each layer weights of depth convolutional neural networks are initial
Change：
Input layer Ι：Input data is the pole view x that a size is (32 × 32)(i)；
Convolutional layer C：Feature of 4 convolutional layers successively to pole view learns, the two dimensional character that each convolutional layer is obtained
The quantity of figure is Ft=(6,8,10,12), the two dimensional character figure of each convolutional layer is obtained according to below equation：
WhereinC (t) q-th of two dimensional character figure is represented, M represents t-1 layers of feature set of graphs, when t-1 is 0, i.e. table
Show the pole view that two dimensional character figure is input,Represent t-1 layers of p-th of characteristic pattern to q-th of two dimension of t layers of convolutional layer
The convolution kernel of characteristic pattern, initial convolution kernel is used as by the matrix of the generating random number (5 × 5) of [- 1,1].Bias is biasing, initially
Change value is 0.(*) represents the computing of convolution, and f () is the linear activation primitive of amendment；It is available by above-mentioned formula：
Convolutional layer C (1) calculates the two dimensional character figure that 6 sizes are (28 × 28)；
Convolutional layer C (2) calculates the two dimensional character figure that 8 sizes are (24 × 24)；
Convolutional layer C (3) calculates the two dimensional character figure that 10 sizes are (20 × 20)；
Convolutional layer C (4) calculates the two dimensional character figure that 12 sizes are (16 × 16)；
Pond layer P：The two dimensional character figure calculated by following formula to convolutional layer C (4) carries out maximum pondization and handled, and obtains
To the eigenmatrix that 12 sizes are (8 × 8)：
Wherein, PMp(u0,v0) the corresponding coordinate of eigenmatrix that the maximum pondization of two dimensional character figure is handled is represented,It is convolution
The two dimensional character figure that layer C (4) is calculated, max () is the maximum of calculating matrix element；
Full linking layer FC (1)：By PMpColumn vector (64 × 1) is arranged in, one is obtained to the full link of each matrix column vector
Dimensional vector L0(768 × 1), i.e. one-dimensional vector L0(768 × 1) as full linking layer FC (1) input vector；
Full linking layer FC (2)：Setting full linking layer FC (1) has 512 neurons, and full linking layer FC (2) has 128 nerves
Member；And the output vector L of full linking layer FC (1) is calculated by following full linking layer propagation formulas1It is used as full linking layer FC (2)
Input vector, and calculate the output vector L of full linking layer FC (2)2It is used as output layer O input vector：
WhereinForWithNetwork weight, is made by the matrix of the generating random number (512 × 768) of [- 1,1]
For initial weight W1, generation (128 × 512) matrix be used as initial weight W2；ForBiasing, initial value is 0；f
() is the linear activation primitive of amendment；Take 1 and 2；
Output layer O：Output layer is set to have K neuron, calculation formula is as follows：
Y '=f (W3L2+b2)；
Wherein, y '(i)For the final output of depth convolutional neural networks, i is i-th of model；W3Random number by [- 1,1]
The matrix of generation (K × 128) is used as initial weight；L2For the output vector of full linking layer FC (2)；
Step s104：The data set χ of pole view is trained, it is 1 to set learning rate η, declines calculation at random using gradient
Method, the result y ' of prediction is exported according to depth convolutional neural networks(i)With real class label y(i)∈ { 1,2 ..., K } mistake
Difference, carries out backpropagation, and 20 algorithms of iteration can restrain, and realizes the internal weights of depth convolutional neural networks after the completion of training
It is determined that；
Step s105：The pole that the threedimensional model that need to extract feature is inputted in the depth convolutional neural networks trained is regarded
Scheme x(i), and calculate the characteristic vector L of the 2nd full linking layer FC (2) outputs2, the threedimensional model pole that feature is extracted needed for being is regarded
The depth characteristic of figure.
The pole view of said extracted threedimensional model refers to：
First, three-dimensional point cloud model is pre-processed, calculates the barycenter and yardstick of three-dimensional point cloud model, and by three-dimensional
Point cloud model is moved in rectangular coordinate system and zoomed in and out, and realizes that three-dimensional point cloud model fastens normalization in rectangular co-ordinate；
Secondly, spherical coordinates is transformed into by the three-dimensional point cloud model of scaling by being fastened in rectangular co-ordinate, and obtain three-dimensional point
The direction of each point of cloud model and distance property；
Again, the spherical coordinates of point set is mapped on the location of pixels of pole view, calculates each pixel sampling distance set
Ultimate range, is used as the ray sampled value of Direction interval；
Finally, the ultimate range of each pixel sampling distance set is arranged in two-dimentional sample graph, is that extracted pole is regarded
Figure.
Specifically, the pole view for extracting threedimensional model specifically includes following steps：
Step S201：Three-dimensional point cloud model is inputted, the yardstick of the three-dimensional point cloud model is P={ pi(xi,yi,zi) | i=1,
2,...,N}；
Step S202：Barycenter g (the g of three-dimensional point cloud model are calculated according to the following equationx, gy, gz), pass through obtained barycenter g
(gx, gy, gz) fasten three-dimensional point cloud model translation transformation to rectangular co-ordinate, then chi of the three-dimensional point cloud model in rectangular coordinate system
Spend for pi'=pi- g, i=1,2 ..., N；Three-dimensional point cloud model p after translation transformationi' barycenter be located at rectangular coordinate system original
Point；
Step s203：The zoom factor s of three-dimensional point cloud model is calculated, three-dimensional point cloud model is zoomed into Unit Scale isOn, wherein, zoom factor s is
Step s204：Spherical coordinates Q is transformed into by the three-dimensional point cloud model of scaling by being fastened in rectangular co-ordinate, it is now three-dimensional
Yardstick of the point cloud model on spherical coordinates beConversion formula is as follows：
Wherein θ ∈ [0, π], it is 0 that the elevation angle is born on semiaxis in Z axis；
Step s205：Spherical coordinates Q is mapped on the location of pixels (u, v) of pole view, mapping relations areThen three-dimensional point cloud model is calculated according to the following equation on the location of pixels (u, v) of pole view；Wherein,
There is spherical coordinates point, multiple spherical coordinates point on one location of pixels (u, v) or in the absence of there is spherical coordinates point；
Wherein nuAnd nvRespectively pole view is wide and long；
Step s206：Each the sampled distance collection of pixel (u, v) isGo out each picture according to the following formula
The maximum of plain sampled distance collection, to obtain pixel sampling value in the view of pole, and is arranged in two-dimentional collection figure and made as ultimate range
For pole view I；
Above-mentioned depth characteristic extracting method directly carries out feature extraction to the pole view of threedimensional model, is regarded for two-dimentional pole
The information of each position can be perceived with convolution operation by desiring to make money or profit.The characteristic pattern obtained by multilayer process of convolution, in full linking layer
It is middle to extract the strong depth characteristic of discrimination property.
And when extracting pole view, preprocessing process needs to carry out threedimensional model Pan and Zoom conversion, it is ensured that three-dimensional
Model is normalized and standardization in standard scale.Threedimensional model point cloud is converted into spherical coordinate system, beneficial to by spherical coordinates
Point is mapped to the respective pixel position of two-dimentional pole view, by mapping, counts the maximum of the distance set of point on the location of pixels
Value, forms two-dimentional sample graph by maximum sampled value, is the new pole view of threedimensional model.
Above-described embodiment is preferably embodiment, but embodiments of the present invention are not by above-described embodiment of the invention
Limitation, other any Spirit Essences without departing from the present invention and the change made under principle, modification, replacement, combine, simplification,
Equivalent substitute mode is should be, is included within protection scope of the present invention.
Claims (4)
1. a kind of depth characteristic extracting method of threedimensional model, it is characterised in that：
First, the pole view of threedimensional model is extracted, the training input data of depth convolutional neural networks is used as；
Secondly, depth convolutional neural networks are built, and pole view is trained；Wherein, depth convolutional neural networks include
Pole view as training input data input layer, learnt for the feature to pole view and obtain the volume of two dimensional character figure
Lamination, for carrying out aggregate statistics to the two dimensional character figure of diverse location and reducing the pond layer of characteristic dimension, for two dimension
Characteristic pattern carries out arrangement link and forms the full linking layer of one-dimensional vector and for exporting the output layer that classification predicts the outcome；
Again, pole view is input into depth convolutional neural networks to be trained, until the convergence of depth convolutional neural networks, realized
The determination of the internal weights of depth convolutional neural networks after the completion of training；
Finally, input need to extract the pole view of the threedimensional model of feature in the depth convolutional neural networks trained, calculate
The characteristic vector of full linking layer in depth convolutional neural networks, then as the depth characteristic for the threedimensional model that need to extract feature.
2. the depth characteristic extracting method of threedimensional model according to claim 1, it is characterised in that：Comprise the following steps：
Step s101：The pole view of threedimensional model is extracted, as the training input data of depth convolutional neural networks, wherein training
Input data is set to x(i)∈ χ, χ are the pole view for having N number of threedimensional model；The corresponding class label of i-th of model is y(i)
∈ { 1,2 ..., K }, K are the categorical measure of threedimensional model；
Step s102：Build depth convolutional neural networks；In depth convolutional neural networks, depth convolutional neural networks include：
Pole view x(i)As the input layer Ι, 4 convolutional layer C (t), t=1,2 of training input data, 3,4,1 pond layer P, two entirely
Linking layer FC (1) and FC (2), and output layer Ο；Wherein, 4 convolutional layer C (t), t=1,2, each convolutional layer and two in 3,4
Each linking layer uses linear activation primitive f (a)=max of amendment (0, a) replacement entirely in individual full linking layer FC (1) and FC (2)
Sigmoid function f (a)=1/ (1+e in depth convolutional neural networks-a)；
Step s103：The parameter of depth convolutional neural networks is set, i.e., by each layer weight initialization of depth convolutional neural networks：
Input layer Ι：Input data is the pole view x that a size is (32 × 32)(i)；
Convolutional layer C：Feature of 4 convolutional layers successively to pole view learns, the two dimensional character figure that each convolutional layer is obtained
Quantity is Ft=(6,8,10,12), the two dimensional character figure of each convolutional layer is obtained according to below equation：
<mrow>
<msubsup>
<mi>x</mi>
<mi>q</mi>
<mi>t</mi>
</msubsup>
<mo>=</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<munder>
<mi>&Sigma;</mi>
<mrow>
<mi>x</mi>
<mo>&Element;</mo>
<mi>M</mi>
</mrow>
</munder>
<msubsup>
<mi>x</mi>
<mi>p</mi>
<mrow>
<mi>t</mi>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msubsup>
<mo>*</mo>
<msubsup>
<mi>k</mi>
<mrow>
<mi>p</mi>
<mi>q</mi>
</mrow>
<mi>t</mi>
</msubsup>
<mo>+</mo>
<msubsup>
<mi>bias</mi>
<mi>q</mi>
<mi>t</mi>
</msubsup>
<mo>)</mo>
</mrow>
</mrow>
WhereinC (t) q-th of two dimensional character figure is represented, M represents t-1 layers of feature set of graphs, when t-1 is 0, that is, represents two dimension
Characteristic pattern is the pole view of input,Represent t-1 layers of p-th of characteristic pattern to q-th of two dimensional character of t layers of convolutional layer
The convolution kernel of figure, initial convolution kernel is used as by the matrix of the generating random number (5 × 5) of [- 1,1].Bias is biasing, initialization value
For 0.(*) represents the computing of convolution, and f () is the linear activation primitive of amendment；It is available by above-mentioned formula：
Convolutional layer C (1) calculates the two dimensional character figure that 6 sizes are (28 × 28)；
Convolutional layer C (2) calculates the two dimensional character figure that 8 sizes are (24 × 24)；
Convolutional layer C (3) calculates the two dimensional character figure that 10 sizes are (20 × 20)；
Convolutional layer C (4) calculates the two dimensional character figure that 12 sizes are (16 × 16)；
Pond layer P：The two dimensional character figure calculated by following formula to convolutional layer C (4) carries out maximum pondization and handled, and obtains 12
Individual size is the eigenmatrix of (8 × 8)：
<mrow>
<msub>
<mi>PM</mi>
<mi>p</mi>
</msub>
<mrow>
<mo>(</mo>
<mrow>
<msub>
<mi>u</mi>
<mn>0</mn>
</msub>
<mo>,</mo>
<msub>
<mi>v</mi>
<mn>0</mn>
</msub>
</mrow>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mi>max</mi>
<mfenced open = "(" close = ")">
<mtable>
<mtr>
<mtd>
<mrow>
<msubsup>
<mi>x</mi>
<mi>p</mi>
<mn>4</mn>
</msubsup>
<mrow>
<mo>(</mo>
<mrow>
<mn>2</mn>
<msub>
<mi>u</mi>
<mn>0</mn>
</msub>
<mo>-</mo>
<mn>1</mn>
<mo>,</mo>
<mn>2</mn>
<msub>
<mi>v</mi>
<mn>0</mn>
</msub>
<mo>-</mo>
<mn>1</mn>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
<mtd>
<mrow>
<msubsup>
<mi>x</mi>
<mi>p</mi>
<mn>4</mn>
</msubsup>
<mrow>
<mo>(</mo>
<mrow>
<mn>2</mn>
<msub>
<mi>u</mi>
<mn>0</mn>
</msub>
<mo>-</mo>
<mn>1</mn>
<mo>,</mo>
<mn>2</mn>
<msub>
<mi>v</mi>
<mn>0</mn>
</msub>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<msubsup>
<mi>x</mi>
<mi>p</mi>
<mn>4</mn>
</msubsup>
<mrow>
<mo>(</mo>
<mrow>
<mn>2</mn>
<msub>
<mi>u</mi>
<mn>0</mn>
</msub>
<mo>,</mo>
<mn>2</mn>
<msub>
<mi>v</mi>
<mn>0</mn>
</msub>
<mo>-</mo>
<mn>1</mn>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
<mtd>
<mrow>
<msubsup>
<mi>x</mi>
<mi>p</mi>
<mn>4</mn>
</msubsup>
<mrow>
<mo>(</mo>
<mrow>
<mn>2</mn>
<msub>
<mi>u</mi>
<mn>0</mn>
</msub>
<mo>,</mo>
<mn>2</mn>
<msub>
<mi>v</mi>
<mn>0</mn>
</msub>
</mrow>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>,</mo>
<mi>p</mi>
<mo>=</mo>
<mn>1</mn>
<mo>,</mo>
<mn>2</mn>
<mo>,</mo>
<mn>...</mn>
<mo>,</mo>
<msup>
<mi>F</mi>
<mn>4</mn>
</msup>
<mo>;</mo>
</mrow>
Wherein, PMp(u0,v0) the corresponding coordinate of eigenmatrix that the maximum pondization of two dimensional character figure is handled is represented,It is convolutional layer C
(4) the two dimensional character figure calculated, max () is the maximum of calculating matrix element；
Full linking layer FC (1)：By PMpBe arranged in column vector (64 × 1), the full link of each matrix column vector is obtained it is one-dimensional to
Measure L0(768 × 1), i.e. one-dimensional vector L0(768 × 1) as full linking layer FC (1) input vector；
Full linking layer FC (2)：Setting full linking layer FC (1) has 512 neurons, and full linking layer FC (2) has 128 neurons；
And the output vector L of full linking layer FC (1) is calculated by following full linking layer propagation formulas1It is used as the defeated of full linking layer FC (2)
Incoming vector, and calculate the output vector L of full linking layer FC (2)2It is used as output layer O input vector：
Ll=f (WlLl-1+bl-1)
Wherein WlFor FC (l-1) and FC (l) network weights, by [- 1,1] generating random number (512 × 768) matrix as first
Beginning weights W1, generation (128 × 512) matrix be used as initial weight W2；bl-1For FC (l-1) biasing, initial value is 0；f(·)
To correct linear activation primitive；L takes 1 and 2；
Output layer O：Output layer is set to have K neuron, calculation formula is as follows：
Y '=f (W3L2+b2)；
Wherein, y '(i)For the final output of depth convolutional neural networks, i is i-th of model；W3Generating random number by [- 1,1]
The matrix of (K × 128) is used as initial weight；L2For the output vector of full linking layer FC (2)；
Step s104：The data set χ of pole view is trained, it is 1 to set learning rate η, uses the random descent algorithm of gradient, root
The result y ' of prediction is exported according to depth convolutional neural networks(i)With real class label y(i)∈ { 1,2 ..., K } error, enters
Row backpropagation, 20 algorithms of iteration can restrain, and realize the determination of the internal weights of depth convolutional neural networks after the completion of training；
Step s105：Input need to extract the pole view x of the threedimensional model of feature in the depth convolutional neural networks trained(i), and calculate the characteristic vector L of the 2nd full linking layer FC (2) outputs2, the threedimensional model pole view of feature is extracted needed for being
Depth characteristic.
3. the depth characteristic extracting method of threedimensional model according to claim 1, it is characterised in that：It is described to extract three-dimensional mould
The pole view of type refers to：
First, three-dimensional point cloud model is pre-processed, calculates the barycenter and yardstick of three-dimensional point cloud model, and by three-dimensional point cloud
Model is moved in rectangular coordinate system and zoomed in and out, and realizes that three-dimensional point cloud model fastens normalization in rectangular co-ordinate；
Secondly, spherical coordinates is transformed into by the three-dimensional point cloud model of scaling by being fastened in rectangular co-ordinate, and obtain three-dimensional point cloud mould
The direction of each point of type and distance property；
Again, the spherical coordinates of point set is mapped on the location of pixels of pole view, calculates the maximum of each pixel sampling distance set
Distance, is used as the ray sampled value of Direction interval；
Finally, the ultimate range of each pixel sampling distance set is arranged in two-dimentional sample graph, is extracted pole view.
4. the depth characteristic extracting method of threedimensional model according to claim 3, it is characterised in that：It is described to extract three-dimensional mould
The pole view of type specifically includes following steps：
Step S201：Three-dimensional point cloud model is inputted, the yardstick of the three-dimensional point cloud model is P={ pi(xi,yi,zi) | i=1,
2,...,N}；
Step S202：Barycenter g (the g of three-dimensional point cloud model are calculated according to the following equationx, gy, gz), pass through obtained barycenter g (gx,
gy, gz) fasten three-dimensional point cloud model translation transformation to rectangular co-ordinate, then three-dimensional point cloud model is in the yardstick of rectangular coordinate system
p′i=pi- g, i=1,2 ..., N；Three-dimensional point cloud model p after translation transformationi' barycenter be located at rectangular coordinate system origin；
<mrow>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mrow>
<msub>
<mi>g</mi>
<mi>x</mi>
</msub>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>N</mi>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</munderover>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<msub>
<mi>g</mi>
<mi>y</mi>
</msub>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>N</mi>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</munderover>
<msub>
<mi>y</mi>
<mi>i</mi>
</msub>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<msub>
<mi>g</mi>
<mi>z</mi>
</msub>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>N</mi>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</munderover>
<msub>
<mi>x</mi>
<mi>z</mi>
</msub>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>;</mo>
</mrow>
Step s203：The zoom factor s of three-dimensional point cloud model is calculated, three-dimensional point cloud model is zoomed into Unit Scale is
On, wherein, zoom factor s is
<mrow>
<mi>s</mi>
<mo>=</mo>
<mi>m</mi>
<mi>a</mi>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mo>|</mo>
<mo>|</mo>
<msubsup>
<mi>p</mi>
<mi>i</mi>
<mo>&prime;</mo>
</msubsup>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mn>2</mn>
<mn>2</mn>
</msubsup>
<mo>)</mo>
</mrow>
<mo>;</mo>
</mrow>
Step s204：Spherical coordinates Q is transformed into by the three-dimensional point cloud model of scaling by being fastened in rectangular co-ordinate, now three-dimensional point cloud
Yardstick of the model on spherical coordinates beConversion formula is as follows：
Wherein θ ∈ [0, π], it is 0 that the elevation angle is born on semiaxis in Z axis；
Step s205：Spherical coordinates Q is mapped on the location of pixels (u, v) of pole view, mapping relations are
Then three-dimensional point cloud model is calculated according to the following equation on the location of pixels (u, v) of pole view；Wherein, location of pixels (u,
V) there is spherical coordinates point, multiple spherical coordinates point on or in the absence of there is spherical coordinates point；
Wherein nuAnd nvRespectively pole view is wide and long；
Step s206：Each the sampled distance collection of pixel (u, v) isGo out each pixel sampling according to the following formula
The maximum of distance set is as ultimate range to obtain pixel sampling value in the view of pole, and be arranged in two-dimentional collection figure to regard as pole
Scheme I；
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710148547.XA CN107092859A (en) | 2017-03-14 | 2017-03-14 | A kind of depth characteristic extracting method of threedimensional model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710148547.XA CN107092859A (en) | 2017-03-14 | 2017-03-14 | A kind of depth characteristic extracting method of threedimensional model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107092859A true CN107092859A (en) | 2017-08-25 |
Family
ID=59648576
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710148547.XA Pending CN107092859A (en) | 2017-03-14 | 2017-03-14 | A kind of depth characteristic extracting method of threedimensional model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107092859A (en) |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108171217A (en) * | 2018-01-29 | 2018-06-15 | 深圳市唯特视科技有限公司 | A kind of three-dimension object detection method based on converged network |
| CN108345831A (en) * | 2017-12-28 | 2018-07-31 | 新智数字科技有限公司 | The method, apparatus and electronic equipment of Road image segmentation based on point cloud data |
| CN108427958A (en) * | 2018-02-02 | 2018-08-21 | 哈尔滨工程大学 | Adaptive weight convolutional neural networks underwater sonar image classification method based on deep learning |
| CN108986159A (en) * | 2018-04-25 | 2018-12-11 | 浙江森马服饰股份有限公司 | A kind of method and apparatus that three-dimensional (3 D) manikin is rebuild and measured |
| CN109063753A (en) * | 2018-07-18 | 2018-12-21 | 北方民族大学 | A kind of three-dimensional point cloud model classification method based on convolutional neural networks |
| CN109064549A (en) * | 2018-07-16 | 2018-12-21 | 中南大学 | Index point detection model generation method and mark point detecting method |
| CN109291657A (en) * | 2018-09-11 | 2019-02-01 | 东华大学 | Laser Jet system is identified based on convolutional neural networks space structure part industry Internet of Things |
| CN109410321A (en) * | 2018-10-17 | 2019-03-01 | 大连理工大学 | Three-dimensional rebuilding method based on convolutional neural networks |
| CN109685848A (en) * | 2018-12-14 | 2019-04-26 | 上海交通大学 | A kind of neural network coordinate transformation method of three-dimensional point cloud and three-dimension sensor |
| CN110059608A (en) * | 2019-04-11 | 2019-07-26 | 腾讯科技（深圳）有限公司 | A kind of object detecting method, device, electronic equipment and storage medium |
| CN110097077A (en) * | 2019-03-26 | 2019-08-06 | 深圳市速腾聚创科技有限公司 | Point cloud data classification method, device, computer equipment and storage medium |
| CN110321910A (en) * | 2018-03-29 | 2019-10-11 | 中国科学院深圳先进技术研究院 | Feature extracting method, device and equipment towards cloud |
| CN111709983A (en) * | 2020-06-16 | 2020-09-25 | 天津工业大学 | Bubble flow field three-dimensional reconstruction method based on convolutional neural network and light field image |
| CN111782879A (en) * | 2020-07-06 | 2020-10-16 | Oppo（重庆）智能科技有限公司 | Model training method and device |
| CN113313830A (en) * | 2021-05-24 | 2021-08-27 | 华南理工大学 | Encoding point cloud feature extraction method based on multi-branch graph convolutional neural network |
| CN113313831A (en) * | 2021-05-24 | 2021-08-27 | 华南理工大学 | Three-dimensional model feature extraction method based on polar coordinate graph convolutional neural network |
| CN113643336A (en) * | 2021-07-26 | 2021-11-12 | 之江实验室 | Three-dimensional image rigid matching method based on spherical polar coordinate system deep neural network |
| CN113744237A (en) * | 2021-08-31 | 2021-12-03 | 华中科技大学 | Deep learning-based automatic detection method and system for muck fluidity |
| CN113777571A (en) * | 2021-08-04 | 2021-12-10 | 中山大学 | Unmanned aerial vehicle cluster dynamic directional diagram synthesis method based on deep learning |
| US11967873B2 (en) | 2019-09-23 | 2024-04-23 | Canoo Technologies Inc. | Fractional slot electric motors with coil elements having rectangular cross-sections |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105868706A (en) * | 2016-03-28 | 2016-08-17 | 天津大学 | Method for identifying 3D model based on sparse coding |
-
2017
- 2017-03-14 CN CN201710148547.XA patent/CN107092859A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105868706A (en) * | 2016-03-28 | 2016-08-17 | 天津大学 | Method for identifying 3D model based on sparse coding |
Non-Patent Citations (5)
| Title |
|---|
| BAOGUANG SHI等: "DeepPano: Deep Panoramic Representation for 3-D Shape Recognition", 《IEEE SIGNAL PROCESSING LETTERS》 * |
| 冯毅攀: "基于视图的三维模型检索技术研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
| 曾向阳: "《智能水中目标识别》", 31 March 2016, 国防工业出版社 * |
| 杜卓明等: "3D模型的加权特征点曲率球面调和表达", 《计算机工程与应用》 * |
| 陈雯柏: "《人工神经网络原理与实践》", 31 January 2016, 西安电子科技大学出版社 * |
Cited By (31)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108345831A (en) * | 2017-12-28 | 2018-07-31 | 新智数字科技有限公司 | The method, apparatus and electronic equipment of Road image segmentation based on point cloud data |
| CN108171217A (en) * | 2018-01-29 | 2018-06-15 | 深圳市唯特视科技有限公司 | A kind of three-dimension object detection method based on converged network |
| CN108427958A (en) * | 2018-02-02 | 2018-08-21 | 哈尔滨工程大学 | Adaptive weight convolutional neural networks underwater sonar image classification method based on deep learning |
| CN110321910B (en) * | 2018-03-29 | 2021-05-28 | 中国科学院深圳先进技术研究院 | Point cloud-oriented feature extraction method, device and equipment |
| CN110321910A (en) * | 2018-03-29 | 2019-10-11 | 中国科学院深圳先进技术研究院 | Feature extracting method, device and equipment towards cloud |
| CN108986159A (en) * | 2018-04-25 | 2018-12-11 | 浙江森马服饰股份有限公司 | A kind of method and apparatus that three-dimensional (3 D) manikin is rebuild and measured |
| CN108986159B (en) * | 2018-04-25 | 2021-10-22 | 浙江森马服饰股份有限公司 | Method and equipment for reconstructing and measuring three-dimensional human body model |
| CN109064549A (en) * | 2018-07-16 | 2018-12-21 | 中南大学 | Index point detection model generation method and mark point detecting method |
| CN109063753A (en) * | 2018-07-18 | 2018-12-21 | 北方民族大学 | A kind of three-dimensional point cloud model classification method based on convolutional neural networks |
| CN109063753B (en) * | 2018-07-18 | 2021-09-14 | 北方民族大学 | Three-dimensional point cloud model classification method based on convolutional neural network |
| CN109291657B (en) * | 2018-09-11 | 2020-10-30 | 东华大学 | Convolutional neural network-based aerospace structure industrial Internet of things identification laser coding system |
| CN109291657A (en) * | 2018-09-11 | 2019-02-01 | 东华大学 | Laser Jet system is identified based on convolutional neural networks space structure part industry Internet of Things |
| CN109410321A (en) * | 2018-10-17 | 2019-03-01 | 大连理工大学 | Three-dimensional rebuilding method based on convolutional neural networks |
| CN109410321B (en) * | 2018-10-17 | 2022-09-20 | 大连理工大学 | Three-dimensional reconstruction method based on convolutional neural network |
| CN109685848A (en) * | 2018-12-14 | 2019-04-26 | 上海交通大学 | A kind of neural network coordinate transformation method of three-dimensional point cloud and three-dimension sensor |
| CN109685848B (en) * | 2018-12-14 | 2023-06-09 | 上海交通大学 | Neural network coordinate transformation method of three-dimensional point cloud and three-dimensional sensor |
| CN110097077A (en) * | 2019-03-26 | 2019-08-06 | 深圳市速腾聚创科技有限公司 | Point cloud data classification method, device, computer equipment and storage medium |
| US11915501B2 (en) | 2019-04-11 | 2024-02-27 | Tencent Technology (Shenzhen) Company Limited | Object detection method and apparatus, electronic device, and storage medium |
| CN110059608A (en) * | 2019-04-11 | 2019-07-26 | 腾讯科技（深圳）有限公司 | A kind of object detecting method, device, electronic equipment and storage medium |
| US11967873B2 (en) | 2019-09-23 | 2024-04-23 | Canoo Technologies Inc. | Fractional slot electric motors with coil elements having rectangular cross-sections |
| CN111709983A (en) * | 2020-06-16 | 2020-09-25 | 天津工业大学 | Bubble flow field three-dimensional reconstruction method based on convolutional neural network and light field image |
| CN111782879A (en) * | 2020-07-06 | 2020-10-16 | Oppo（重庆）智能科技有限公司 | Model training method and device |
| CN113313830B (en) * | 2021-05-24 | 2022-12-16 | 华南理工大学 | Encoding point cloud feature extraction method based on multi-branch graph convolutional neural network |
| CN113313831B (en) * | 2021-05-24 | 2022-12-16 | 华南理工大学 | Three-dimensional model feature extraction method based on polar coordinate graph convolution neural network |
| CN113313831A (en) * | 2021-05-24 | 2021-08-27 | 华南理工大学 | Three-dimensional model feature extraction method based on polar coordinate graph convolutional neural network |
| CN113313830A (en) * | 2021-05-24 | 2021-08-27 | 华南理工大学 | Encoding point cloud feature extraction method based on multi-branch graph convolutional neural network |
| CN113643336A (en) * | 2021-07-26 | 2021-11-12 | 之江实验室 | Three-dimensional image rigid matching method based on spherical polar coordinate system deep neural network |
| CN113643336B (en) * | 2021-07-26 | 2024-03-15 | 之江实验室 | Three-dimensional image rigid matching method based on spherical polar coordinate system depth neural network |
| CN113777571A (en) * | 2021-08-04 | 2021-12-10 | 中山大学 | Unmanned aerial vehicle cluster dynamic directional diagram synthesis method based on deep learning |
| CN113777571B (en) * | 2021-08-04 | 2023-08-11 | 中山大学 | Unmanned aerial vehicle cluster dynamic pattern synthesis method based on deep learning |
| CN113744237A (en) * | 2021-08-31 | 2021-12-03 | 华中科技大学 | Deep learning-based automatic detection method and system for muck fluidity |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107092859A (en) | A kind of depth characteristic extracting method of threedimensional model | |
| Yan et al. | Second: Sparsely embedded convolutional detection | |
| Lu et al. | 3DCTN: 3D convolution-transformer network for point cloud classification | |
| Zhang et al. | Latentgnn: Learning efficient non-local relations for visual recognition | |
| Kang et al. | Depth-adaptive deep neural network for semantic segmentation | |
| Wang et al. | An efficient and effective convolutional auto-encoder extreme learning machine network for 3d feature learning | |
| Wang et al. | Cross self-attention network for 3D point cloud | |
| Teow | Understanding convolutional neural networks using a minimal model for handwritten digit recognition | |
| CN111242208A (en) | Point cloud classification method, point cloud segmentation method and related equipment | |
| CN107437096A (en) | Image classification method based on the efficient depth residual error network model of parameter | |
| CN106845499A (en) | A kind of image object detection method semantic based on natural language | |
| CN108334830A (en) | A kind of scene recognition method based on target semanteme and appearance of depth Fusion Features | |
| Jiang et al. | An eight-layer convolutional neural network with stochastic pooling, batch normalization and dropout for fingerspelling recognition of Chinese sign language | |
| Khalifa et al. | Deep galaxy V2: Robust deep convolutional neural networks for galaxy morphology classifications | |
| CN114049381A (en) | Twin cross target tracking method fusing multilayer semantic information | |
| Sun et al. | Vicinity vision transformer | |
| CN105224935A (en) | A kind of real-time face key point localization method based on Android platform | |
| CN114638408B (en) | Pedestrian track prediction method based on space-time information | |
| Tao et al. | Pooling operations in deep learning: from “invariable” to “variable” | |
| CN110197255A (en) | A kind of deformable convolutional network based on deep learning | |
| Yang et al. | Detection of river floating garbage based on improved YOLOv5 | |
| Gao et al. | Natural scene recognition based on convolutional neural networks and deep Boltzmannn machines | |
| Ning et al. | Point-voxel and bird-eye-view representation aggregation network for single stage 3D object detection | |
| Fan et al. | Hcpvf: Hierarchical cascaded point-voxel fusion for 3D object detection | |
| Gao et al. | Joint learning of semantic segmentation and height estimation for remote sensing image leveraging contrastive learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170825 |