CN110223382A - Single-frame images free view-point method for reconstructing three-dimensional model based on deep learning - Google Patents
Single-frame images free view-point method for reconstructing three-dimensional model based on deep learning Download PDFInfo
- Publication number
- CN110223382A CN110223382A CN201910509328.9A CN201910509328A CN110223382A CN 110223382 A CN110223382 A CN 110223382A CN 201910509328 A CN201910509328 A CN 201910509328A CN 110223382 A CN110223382 A CN 110223382A
- Authority
- CN
- China
- Prior art keywords
- point
- model
- camera
- free view
- view
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Geometry (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computer Hardware Design (AREA)
- Computer Graphics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
- Processing Or Creating Images (AREA)
Abstract
The invention discloses a kind of single-frame images free view-point method for reconstructing three-dimensional model based on deep learning, comprising the following steps: training sample generates;Picture high-level semantic is obtained using feature extraction network;The unrelated threedimensional model point cloud of output viewpoint and camera view parameter are converted by image, semantic decoupling by Decoupling network;The unrelated reconstructing three-dimensional model of viewpoint;Camera view estimation is generated with free view-point;Free view-point threedimensional model generates;And deep learning model training.The method of the present invention can simply and efficiently rebuild the threedimensional model of free view-point from single-frame images, improve the generalization of model, widened use scope.
Description
Technical field
The present invention relates to reconstructing three-dimensional model field more particularly to a kind of single-frame images free view-points based on deep learning
Method for reconstructing three-dimensional model.
Background technique
Free view-point threedimensional model is to allow people same when switching viewing angle on the basis of common threedimensional model
Three-dimensional sense having the same, this can provide the visual environment of more true nature for three-dimensional multimedia.Due to three-dimensional mould
The complexity of type itself, conventional method generation free view-point threedimensional model cost is higher, needs worker people under different points of view
Work rendering generates the threedimensional model of more different points of view, and efficiency is lower, complicated for operation.How free view-point is simply and efficiently generated
Threedimensional model always is the research hotspot of researcher, possesses huge application potential.
The unrelated threedimensional model of viewpoint can be regarded as a kind of special, free view-point threedimensional model under initial viewpoint, and two
Mould shapes are identical between person, only there is viewpoint difference.It can be given birth to by carrying out viewpoint change to the unrelated threedimensional model of viewpoint
At the threedimensional model of free view-point.The unrelated threedimensional model of viewpoint has extensively in Attitude estimation, object tracking, the fields such as target detection
General application prospect.Such as in three-dimension object Attitude estimation, researcher need will the unrelated threedimensional model of viewpoint that pre-establish with
Two-dimensional silhouette in image carries out feature adaptation, to realize Attitude estimation;In three-dimension object tracing detection, often there is
The case where significant viewpoint of object changes needs video camera to move object dimensional and carries out the unrelated tracking of viewpoint, at this time convenient for efficient
Ground carries out feature extraction and matches with result.
Currently, deep learning achieves great success in single-frame images three-dimensional reconstruction field.Researcher utilizes convolutional Neural
The powerful ability in feature extraction of network can be easily accomplished to priori knowledges such as the shape semanteme of single-frame images and viewpoint semantemes
It sufficiently extracts, and then obtains the higher level of abstraction semantic feature with powerful generalization ability, one is passed through to higher level of abstraction semantic feature
Fixed mapping becomes the geometric parameter with certain sense, and then instructs reconstructing three-dimensional model.But many researchs are always by three at present
The study of dimension module shape and viewpoint is bound together, and the threedimensional model of generation is only suitable for single camera view, is lacked certain
Changeability can not follow viewpoint to change and flexibly change, more limit in practical applications.How this kind generation method is widened
Use scope, improve the formation efficiency of threedimensional model, be one make us puzzlement the problem of.
Summary of the invention
Technical problem to be solved by the invention is to provide a kind of, and the single-frame images free view-point based on deep learning is three-dimensional
Model reconstruction method can simply and efficiently rebuild the threedimensional model of free view-point from single-frame images.
In order to solve the above technical problems, the technical solution adopted by the present invention is that:
A kind of single-frame images free view-point method for reconstructing three-dimensional model based on deep learning, comprising the following steps:
Step 1: sampling CAD model and rendered, and it is different from different points of view to generate initial viewpoint true shape point cloud
Single-frame images under;
Step 2: by the intensification of feature extraction network, the high-level semantic of image is gradually obtained;
Step 3: being converted image high-level semantic by Decoupling network, the point cloud coordinate of the unrelated threedimensional model of output viewpoint
With camera view parameter;
Step 4: the unrelated threedimensional model point cloud coordinate of the viewpoint that Decoupling network is exported is fitted by amendment with triangular plate
3D shape reconstruction is carried out, the unrelated threedimensional model of viewpoint is obtained;
Step 5: obtaining camera view by homogeneous transformation to the camera view parameter of Decoupling network output, basic herein
On converted, generate free view-point;
Step 6: free view-point threedimensional model unrelated with viewpoint is multiplied, free view-point threedimensional model is obtained;
Step 7: training sample input neural network is trained automatically, and progressive updating network parameter optimizes free view-point three
Dimension module obtains optimal result.
Further, in step 1, sampling is carried out to CAD model using OpenGL and renders generation training sample.
Further, in step 2, using ResNet as feature extraction network, i.e., using following formula to each trained sample
This input picture carries out feature extraction:
Wherein, N is positive integer;Indicate that jth is opened produced by image in i-th of CAD training sample in n-th of classification
Semantic information;ResNet indicates feature extraction network;Indicate that jth is schemed in i-th of CAD training sample in n-th of classification
Picture.
Further, the step 3 specifically:
The high-level semantic that will be extracted is averaged pondization and full articulamentum to the feature after convolution using the neural network overall situation
Coupling conversion is illustrated, the unrelated threedimensional model point cloud of viewpoint and camera view parameter are obtained;The camera view parameter includes: camera posture
It is indicated by Eulerian angles, including pitch angle pitch (γ), roll angle roll (β), angle of drift yaw (α);Camera coordinates are existed by camera
Coordinate t under initial coordinate systemx,ty,tzIt indicates.
Further, the step 4 specifically:
The point cloud coordinate of Decoupling network output is modified, then to the point cloud coordinate of dense distribution in blocks with triangular plate
Fitting forms the unrelated three-dimensional surface model of viewpoint;
The negative value in point cloud coordinate required for the unrelated three-dimensional reconstruction of viewpoint is corrected by following formula:
Wherein, Y indicate output as a result,Indicate final output, ReLU function representation positive value amendment response unit.
Further, the step 5 specifically:
Camera posture and coordinate are obtained by homogeneous transformation to the camera view parameter of Decoupling network output;By three kinds of Eulers
Spin matrix is calculated in angle, including pitch angle pitch (γ), roll angle roll (β), angle of drift yaw (α), indicates camera appearance
State;tx,ty,tzCamera is represented under initial coordinate system away from the distance of CAD model, indicates camera coordinates;It is obtained by homogeneous transformation
Coordinate of the CAD model under camera coordinates system, such as following formula:
T=(tx,ty,tz)T
Wherein, x, y, z represent the fixed coordinates of CAD model;X', y', z' represent the seat of the CAD model under camera coordinates system
Mark;R is spin matrix, includes that pitch angle pitch (γ), roll angle roll (β) and angle of drift yaw (α) are calculated by Eulerian angles
It arrives, indicates camera posture;tx,ty,tzCamera is represented under initial coordinate system away from the distance of CAD model, indicates camera coordinates;T is
Posture changing matrix represent camera view, including posture and coordinate;
On the basis of the camera view of estimation, free view-point is generated;Using CAD model as the centre of sphere, change camera coordinates in ball
It is moved on face, by camera pose adjustment to towards CAD model is directed at, records R' at this time, t' obtains free view-point, as follows
Formula:
T'=(x, y, z)
Wherein, t' is the coordinate that (x, y, z) represents free view-point, and R' represents camera posture at this time, and T' is indicated at this time
Posture changing matrix represent the posture and coordinate of free view-point.
Further, the step 6 are as follows:
The unrelated threedimensional model of the viewpoint learnt is multiplied with free view-point by following formula, obtains the three-dimensional mould of free view-point
Type:
Wherein, ModeliIndicate the unrelated threedimensional model of viewpoint, ModelcIndicate free view-point threedimensional model.
Further, the step 7 specifically:
Pass through the chamfering distance and earth movement between the threedimensional model and true free view-point threedimensional model of neural network forecast
The weighted sum of distance is to deep learning model training;Such as following formula:
Loss=λ1lossEMD+λ2lossCD
Wherein, lossEMDWith lossCDRespectively represent neural network forecast threedimensional model and true free view-point threedimensional model it
Between chamfering distance and earth moving distance lose;λ1, λ2Represent loss weight;P represents the threedimensional model of neural network forecast;Q is represented
True free view-point threedimensional model;| | | | 2 indicate two norms;F (x) represents equivalent mappings.
Compared with prior art, the beneficial effects of the present invention are: 1) being based on deep learning neural network, training study is completed
Three-dimensional reconstruction task improves the efficiency of three-dimensional reconstruction, reduces operation difficulty;2) neural network decoupling method is utilized, by three
Dimension module shape similarity metric and viewpoint study decouple, and can complete the unrelated threedimensional model of viewpoint and camera view is estimated, powerful;
3) camera view on this basis, learnt can be transformed to free view-point, and threedimensional model unrelated with viewpoint is multiplied to obtain certainly
By the threedimensional model under viewpoint, application range is wider.
Detailed description of the invention
Fig. 1 is the flow chart of the single-frame images free view-point method for reconstructing three-dimensional model based on deep learning;
Fig. 2 is that training sample generates schematic diagram;
Fig. 3 is the data transmission scheme of feature extraction network;
Fig. 4 is image, semantic decoupling figure;
Fig. 5 is free view-point reconstructing three-dimensional model process schematic.
Specific embodiment
The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.The present invention passes through depth
Neural network is practised, the unrelated reconstructing three-dimensional model of viewpoint in three-dimensional reconstruction is decoupled with camera view estimation task.On the one hand, depending on
The unrelated threedimensional model of point does not change with the variation of camera view, possesses significant migration and suitability;On the other hand, may be used
Free view-point is derived by the camera view estimated, threedimensional model unrelated with viewpoint is multiplied, free view-point threedimensional model is obtained,
Expand the use scope of this method.
It is worth noting that, neural network decoupling of the invention does not use any compulsory measure, but pass through viewpoint
Independence model and free view-point generate this physical computing of free view-point threedimensional model jointly voluntarily to guide completion, this is effectively
To demonstrate neural network itself can be guided and specification study by physical computing, not needed using supervising by force
Mode carries out the study of black box formula, has certain interpretation and stronger convincingness.
The present invention can realize that programming language is also alternatively, can use on Windows and Linux platform
Python is realized.
As shown in Figure 1, the single-frame images free view-point method for reconstructing three-dimensional model based on deep learning includes the following steps:
Step 1: training sample is generated
Sampling is carried out to CAD model using OpenGL and renders generation training sample, is used for neural metwork training.Certainly,
The method for generating training sample is not limited to OpenGL, and any method that can be realized identical technical effect may be used to generate
Training sample.
Select CAD data collection.Using Princeton provide ModelNet data set, the data set computer vision,
Content is comprehensive, and model object is excellent using extensively, covering for computer graphics, robot and cognitive science field.ModelNet
Target classification in a total of 662, the data of 127915 CAD and the labeled direction direction of ten classes include three subsets.This
Embodiment selects ModelNet40 subset, which includes the CAD model of 40 classifications.
Single-frame images generates.Using the camera analog functuion of OpenGL, three-dimension object is put into appropriate location in the scene,
Camera is adjusted to different angle and different distance, is projected in three-dimension object on two-dimentional film, the fixed view for using 224*224
Mouthful, the image after being rinsed, each CAD model generates 24 images, while record the corresponding camera angle of every image and
Distance, as shown in the right side of fig 2.
The sampling of CAD model true shape.The present embodiment utilizes OpenGL, CAD model is secured in place, to CAD
Model carries out Points Sample, obtains model true shape, acquires 4096 points altogether, as shown in the left side Fig. 2.
Training set and test set divide.In order to facilitate training, the present embodiment upsets obtained sample data set, every one kind
4/5 training dataset Chou Qu not be collectively constituted;Remaining 1/5 data independent storage of all categories, for testing the mould of each type
Type effect.
Step 2: image high-level semantic extracts
Feature extraction is carried out to the input picture of each training sample, obtains the high-level semantics features of input picture.Due to
Each CAD model has 24 subtended angle degree to be made respectively from apart from different input pictures, therefore to 24 images of each sample
The extraction of high-level semantic is carried out with identical Feature Selection Model.The present embodiment using ResNet as feature extraction network, when
So, the feature extraction network of other numbers of plies can also be selected in other examples.It is directed to the three of differing complexity simultaneously
Dimension module, using the network of different depth, including ResNe51With ResNe101。
Feature extraction can be carried out by input picture of the following formula to each training sample:
Wherein, N is positive integer, and N takes 40 in the present embodiment, indicates 40 training data classes in ModelNet40 subset
Not;Indicate that jth opens semantic information caused by image, the maximum value of j in i-th of CAD training sample in n-th of classification
It is 24;ResNet indicates feature extraction network;Indicate that jth opens image in i-th of CAD training sample in n-th of classification.
ResNet () core is short circuit connection (shortcut) and structure block (building block).All features
It extracts network and is all divided into 5 parts, composition situation is respectively: conv1 is common convolutional layer;Conv2_x, conv3_x, conv4_
X, conv5_x are structure block (building block), connect (shortcut) with short circuit by convolutional layer and form, on the left of Fig. 4
It is shown.
Short circuit connection (shortcut) participate in composition structure block (building block) when, for input feature vector figure with it is defeated
Whether characteristic pattern channel number is identical out, to be divided into two kinds of situations and consider.It inputs the following formula of relationship between output:
Wherein, x represents input feature vector figure, and F (x) represents output characteristic pattern, and c represents the number of active lanes of characteristic pattern, and W is convolution
Operation, for adjusting the number of active lanes of characteristic pattern.
Step 3: image high-level semantic decoupling
The high-level semantic that will be extracted is averaged pondization and full articulamentum to the feature after convolution using the neural network overall situation
Coupling conversion is illustrated, the unrelated threedimensional model point cloud of viewpoint and camera view parameter are obtained.The camera view parameter includes: camera appearance
State is indicated by Eulerian angles (pitch angle pitch (γ), roll angle roll (β), angle of drift yaw (α));Camera position is existed by camera
Coordinate t under initial coordinate systemx,ty,tz(CAD model is origin) indicates;
In the average Chi Huazhong of the overall situation, the characteristic pattern overall situation in a channel be averaged Chi Huahou obtain one it is global it is semantic averagely
Value, i.e. a single element output.The following formula of relationship between the input and output of the average pond layer of the overall situation:
GAP (i)=mean (conv (i))
Wherein, GAP (i) is global average pond layer output, and conv (i) is input feature vector figure, and mean () representative takes input
The average value of data global area, i represent input and output channel number;
In full articulamentum, the output in global average pond is inputted as full articulamentum, by the way that full articulamentum is arranged
Neuron number, the point cloud coordinate of matching free view-point threedimensional model and the total quantity of camera view estimation parameter.Full articulamentum
Input with output the following formula of relationship:
Wherein, X represents input data, and i indicates input channel number;Y expression output data, j expression output channel number, i.e., three
The total quantity of dimension module point cloud and camera view parameter.
Step 4: the unrelated reconstructing three-dimensional model of viewpoint
The point cloud coordinate of Decoupling network output is modified, it is then quasi- with triangular plate to the point cloud of dense distribution in blocks
It closes, forms the unrelated three-dimensional surface model of the viewpoint of continuous, accurate, good state.
The negative value in point cloud coordinate required for the unrelated reconstruction of viewpoint is corrected by following formula:
Wherein, Y indicate output as a result,Indicate final output, ReLU function representation positive value amendment response unit.
Step 5: camera view estimation is generated with free view-point
Camera posture and coordinate are obtained by homogeneous transformation to the camera view parameter of Decoupling network output.By three kinds of Eulers
Spin matrix is calculated in angle, including pitch angle pitch (γ), roll angle roll (β), angle of drift yaw (α), indicates camera appearance
State;tx,ty,tzCamera is represented under initial coordinate system away from the distance of CAD model, indicates camera coordinates;It can be obtained by homogeneous transformation
To coordinate of the CAD model under camera coordinates system, such as following formula:
T=(tx,ty,tz)T
Wherein, x, y, z represent the fixation position of CAD model;X', y', z' represent the seat of the CAD model under camera coordinates system
Mark;R is spin matrix, includes that pitch angle pitch (γ), roll angle roll (β) and angle of drift yaw (α) are calculated by Eulerian angles
It arrives, indicates camera posture;tx,ty,tzCamera is represented under initial coordinate system away from the distance of CAD model, indicates camera coordinates;T is
Posture changing matrix represent camera view (including posture and coordinate).
On the basis of the camera view of estimation, free view-point is generated.Using CAD model as the centre of sphere, change camera position in ball
It is moved on face, by camera pose adjustment to towards CAD model is directed at, records R' at this time, free view-point can be obtained in t', such as
Following formula:
T'=(x, y, z)
Wherein, t' is the position that (x, y, z) represents free view-point, and R' represents camera posture at this time, and T' is indicated at this time
Posture changing matrix represent free view-point (including posture and coordinate).
Step 6: free view-point threedimensional model generates
The unrelated threedimensional model of viewpoint and free view-point threedimensional model only exist the gap of viewpoint, i.e. camera posture and position
Difference, but the shape information of threedimensional model is identical.It is regarded by following formula by the unrelated threedimensional model of the viewpoint learnt and freely
Point is multiplied, and the threedimensional model of free view-point can be obtained:
Wherein, ModeliIndicate the unrelated threedimensional model of viewpoint, ModelcIndicate free view-point threedimensional model.
Step 7: deep learning model training
Pass through the chamfering distance and earth movement between the threedimensional model and true free view-point threedimensional model of neural network forecast
The weighted sum of distance is to deep learning model training;Such as following formula:
Loss=λ1lossEMD+λ2lossCD
Wherein, lossEMDWith lossCDRespectively represent neural network forecast threedimensional model and true free view-point threedimensional model it
Between chamfering distance and earth moving distance lose;λ1, λ2Represent loss weight;P represents the threedimensional model of neural network forecast;Q is represented
True free view-point threedimensional model;| | | | 2 indicate two norms;F (x) represents equivalent mappings.
Single-frame images free view-point method for reconstructing three-dimensional model provided by the invention based on deep learning is different from previous
Three-dimensional rebuilding method, the present invention can without manual operations, restore threedimensional model from single image automatically.It is more unique
It is that the present invention separates the generation of the unrelated threedimensional model of viewpoint with camera view estimation, while generating free view-point three-dimensional mould
Type.The unrelated threedimensional model of viewpoint can be Attitude estimation, and the fields such as tracing detection provide convenience;Free view-point threedimensional model is available
In enlarging 3-D data set, data cost is reduced, promotes three-dimensional reconstruction working efficiency.Generally, opposite conventional operation, the present invention
It provides the more flexible method for reconstructing three-dimensional model of one kind and improves the general of model under the premise of guaranteeing basic reconstruction tasks
The property changed, has widened use scope.
Claims (8)
1. a kind of single-frame images free view-point method for reconstructing three-dimensional model based on deep learning, which is characterized in that including following
Step:
Step 1: sampling CAD model and rendered, and initial viewpoint true shape point cloud and different points of view different distance are generated
Under single-frame images;
Step 2: by the intensification of feature extraction network, the high-level semantic of image is gradually obtained;
Step 3: being converted image high-level semantic by Decoupling network, the point cloud coordinate and phase of the unrelated threedimensional model of output viewpoint
Machine viewpoint parameter;
Step 4: the unrelated threedimensional model point cloud coordinate of the viewpoint that Decoupling network is exported is fitted with triangular plate and is carried out by amendment
3D shape is rebuild, and the unrelated threedimensional model of viewpoint is obtained;
Step 5: obtaining camera view by homogeneous transformation to the camera view parameter of Decoupling network output, on this basis into
Row transformation, generates free view-point;
Step 6: free view-point threedimensional model unrelated with viewpoint is multiplied, free view-point threedimensional model is obtained;
Step 7: training sample input neural network is trained automatically, and progressive updating network parameter optimizes free view-point three-dimensional mould
Type obtains optimal result.
2. the single-frame images free view-point method for reconstructing three-dimensional model based on deep learning as described in claim 1, feature
It is, in step 1, sampling is carried out to CAD model using OpenGL and renders generation training sample.
3. the single-frame images free view-point method for reconstructing three-dimensional model based on deep learning as described in claim 1, feature
It is, in step 2, using ResNet as feature extraction network, i.e., using following formula to the input picture of each training sample
Carry out feature extraction:
Wherein, N is positive integer;Indicate that jth opens language caused by image in i-th of CAD training sample in n-th of classification
Adopted information;ResNet indicates feature extraction network;Indicate that jth opens image in i-th of CAD training sample in n-th of classification.
4. the single-frame images free view-point method for reconstructing three-dimensional model based on deep learning as described in claim 1, feature
It is, the step 3 specifically:
The high-level semantic that will be extracted is averaged pondization and full articulamentum to the feature diagram after convolution using the neural network overall situation
Coupling conversion, obtains the unrelated threedimensional model point cloud of viewpoint and camera view parameter;The camera view parameter includes: camera posture by Europe
Angle, including pitch angle pitch (γ), roll angle roll (β), angle of drift yaw (α) is drawn to indicate;Camera coordinates are by camera initial
Coordinate t under coordinate systemx,ty,tzIt indicates.
5. the single-frame images free view-point method for reconstructing three-dimensional model based on deep learning as described in claim 1, feature
It is, the step 4 specifically:
The point cloud coordinate of Decoupling network output is modified, it is then quasi- with triangular plate to the point cloud coordinate of dense distribution in blocks
It closes, forms the unrelated three-dimensional surface model of viewpoint;
The negative value in point cloud coordinate required for the unrelated three-dimensional reconstruction of viewpoint is corrected by following formula:
Wherein, Y indicate output as a result,Indicate final output, ReLU function representation positive value amendment response unit.
6. the single-frame images free view-point method for reconstructing three-dimensional model based on deep learning as described in claim 1, feature
It is, the step 5 specifically:
Camera posture and coordinate are obtained by homogeneous transformation to the camera view parameter of Decoupling network output;By three kinds of Eulerian angles,
Including pitch angle pitch (γ), roll angle roll (β), angle of drift yaw (α), spin matrix is calculated, indicates camera posture;
tx,ty,tzCamera is represented under initial coordinate system away from the distance of CAD model, indicates camera coordinates;CAD is obtained by homogeneous transformation
Coordinate of the model under camera coordinates system, such as following formula:
T=(tx,ty,tz)T
Wherein, x, y, z represent the fixed coordinates of CAD model;X', y', z' represent the CAD model coordinate under camera coordinates system;R is
Spin matrix includes that pitch angle pitch (γ), roll angle roll (β) and angle of drift yaw (α) are calculated by Eulerian angles, is indicated
Camera posture;tx,ty,tzCamera is represented under initial coordinate system away from the distance of CAD model, indicates camera coordinates;T is that posture becomes
Matrix is changed, camera view, including posture and coordinate are represented;
On the basis of the camera view of estimation, free view-point is generated;Using CAD model as the centre of sphere, change camera coordinates on spherical surface
Movement records R' at this time by camera pose adjustment to towards CAD model is directed at, and t' obtains free view-point, such as following formula:
T'=(x, y, z)
Wherein, t' is the coordinate that (x, y, z) represents free view-point, and R' represents camera posture at this time, and T' indicates posture at this time
Transformation matrix represents the posture and coordinate of free view-point.
7. the single-frame images free view-point method for reconstructing three-dimensional model based on deep learning as described in claim 1, feature
It is, the step 6 are as follows:
The unrelated threedimensional model of the viewpoint learnt is multiplied with free view-point by following formula, obtains the threedimensional model of free view-point:
Wherein, ModeliIndicate the unrelated threedimensional model of viewpoint, ModelcIndicate free view-point threedimensional model.
8. the single-frame images free view-point method for reconstructing three-dimensional model based on deep learning as described in claim 1, feature
It is, the step 7 specifically:
Pass through the chamfering distance and earth moving distance between the threedimensional model and true free view-point threedimensional model of neural network forecast
Weighted sum to deep learning model training;Such as following formula:
Loss=λ1lossEMD+λ2lossCD
Wherein, lossEMDWith lossCDIt respectively represents between the threedimensional model of neural network forecast and true free view-point threedimensional model
Chamfering distance is lost with earth moving distance;λ1, λ2Represent loss weight;P represents the threedimensional model of neural network forecast;Q represents true
Free view-point threedimensional model;| | | | 2 indicate two norms;F (x) represents equivalent mappings.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910509328.9A CN110223382B (en) | 2019-06-13 | 2019-06-13 | Single-frame image free viewpoint three-dimensional model reconstruction method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910509328.9A CN110223382B (en) | 2019-06-13 | 2019-06-13 | Single-frame image free viewpoint three-dimensional model reconstruction method based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110223382A true CN110223382A (en) | 2019-09-10 |
CN110223382B CN110223382B (en) | 2021-02-12 |
Family
ID=67816761
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910509328.9A Active CN110223382B (en) | 2019-06-13 | 2019-06-13 | Single-frame image free viewpoint three-dimensional model reconstruction method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110223382B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110675488A (en) * | 2019-09-24 | 2020-01-10 | 电子科技大学 | Construction method of creative three-dimensional voxel model modeling system based on deep learning |
CN110798673A (en) * | 2019-11-13 | 2020-02-14 | 南京大学 | Free viewpoint video generation and interaction method based on deep convolutional neural network |
CN113204988A (en) * | 2019-02-05 | 2021-08-03 | 辉达公司 | Small sample viewpoint estimation |
CN113569761A (en) * | 2021-07-30 | 2021-10-29 | 广西师范大学 | Student viewpoint estimation method based on deep learning |
CN115375884A (en) * | 2022-08-03 | 2022-11-22 | 北京微视威信息科技有限公司 | Free viewpoint synthesis model generation method, image rendering method and electronic device |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101216896A (en) * | 2008-01-14 | 2008-07-09 | 浙江大学 | An identification method for movement by human bodies irrelevant with the viewpoint based on stencil matching |
CN101719286A (en) * | 2009-12-09 | 2010-06-02 | 北京大学 | Multiple viewpoints three-dimensional scene reconstructing method fusing single viewpoint scenario analysis and system thereof |
JP5295416B1 (en) * | 2012-08-01 | 2013-09-18 | ヤフー株式会社 | Image processing apparatus, image processing method, and image processing program |
US20130321393A1 (en) * | 2012-05-31 | 2013-12-05 | Microsoft Corporation | Smoothing and robust normal estimation for 3d point clouds |
CN106778687A (en) * | 2017-01-16 | 2017-05-31 | 大连理工大学 | Method for viewing points detecting based on local evaluation and global optimization |
US20180152688A1 (en) * | 2016-11-28 | 2018-05-31 | Sony Corporation | Decoder-centric uv codec for free-viewpont video streaming |
US20180220125A1 (en) * | 2017-01-31 | 2018-08-02 | Tetavi Ltd. | System and method for rendering free viewpoint video for sport applications |
-
2019
- 2019-06-13 CN CN201910509328.9A patent/CN110223382B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101216896A (en) * | 2008-01-14 | 2008-07-09 | 浙江大学 | An identification method for movement by human bodies irrelevant with the viewpoint based on stencil matching |
CN101719286A (en) * | 2009-12-09 | 2010-06-02 | 北京大学 | Multiple viewpoints three-dimensional scene reconstructing method fusing single viewpoint scenario analysis and system thereof |
US20130321393A1 (en) * | 2012-05-31 | 2013-12-05 | Microsoft Corporation | Smoothing and robust normal estimation for 3d point clouds |
JP5295416B1 (en) * | 2012-08-01 | 2013-09-18 | ヤフー株式会社 | Image processing apparatus, image processing method, and image processing program |
US20180152688A1 (en) * | 2016-11-28 | 2018-05-31 | Sony Corporation | Decoder-centric uv codec for free-viewpont video streaming |
CN106778687A (en) * | 2017-01-16 | 2017-05-31 | 大连理工大学 | Method for viewing points detecting based on local evaluation and global optimization |
US20180220125A1 (en) * | 2017-01-31 | 2018-08-02 | Tetavi Ltd. | System and method for rendering free viewpoint video for sport applications |
Non-Patent Citations (5)
Title |
---|
JIAGENG FENG等: "View-invariant human action recognition via robust locally adaptive multi-view learning", 《FITEE》 * |
JIANMEI SU等: "Robust spatial–temporal Bayesian view synthesis for video stitching with occlusion handling", 《MACHINE VISION AND APPLICATIONS》 * |
人工智能博士: "中科院自动化所--模式识别实验室-申抒含-《基于图像的大规模场景三维建模——从几何重建到语义矢量重建》学习记录", 《HTTPS://BLOG.CSDN.NET/QQ_15698613/ARTICLE/DETAILS/89406704》 * |
冯家更 等: "视点无关的行为识别综述", 《中国图像图形学报》 * |
杜军辉: "基于集成成像光场信息的三维计算重建方法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113204988A (en) * | 2019-02-05 | 2021-08-03 | 辉达公司 | Small sample viewpoint estimation |
CN110675488A (en) * | 2019-09-24 | 2020-01-10 | 电子科技大学 | Construction method of creative three-dimensional voxel model modeling system based on deep learning |
CN110675488B (en) * | 2019-09-24 | 2023-02-28 | 电子科技大学 | Method for constructing modeling system of creative three-dimensional voxel model based on deep learning |
CN110798673A (en) * | 2019-11-13 | 2020-02-14 | 南京大学 | Free viewpoint video generation and interaction method based on deep convolutional neural network |
CN110798673B (en) * | 2019-11-13 | 2021-03-19 | 南京大学 | Free viewpoint video generation and interaction method based on deep convolutional neural network |
CN113569761A (en) * | 2021-07-30 | 2021-10-29 | 广西师范大学 | Student viewpoint estimation method based on deep learning |
CN113569761B (en) * | 2021-07-30 | 2023-10-27 | 广西师范大学 | Student viewpoint estimation method based on deep learning |
CN115375884A (en) * | 2022-08-03 | 2022-11-22 | 北京微视威信息科技有限公司 | Free viewpoint synthesis model generation method, image rendering method and electronic device |
CN115375884B (en) * | 2022-08-03 | 2023-05-30 | 北京微视威信息科技有限公司 | Free viewpoint synthesis model generation method, image drawing method and electronic device |
Also Published As
Publication number | Publication date |
---|---|
CN110223382B (en) | 2021-02-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110223382A (en) | Single-frame images free view-point method for reconstructing three-dimensional model based on deep learning | |
CN111339903B (en) | Multi-person human body posture estimation method | |
WO2022121645A1 (en) | Method for generating sense of reality of virtual object in teaching scene | |
CN109214282B (en) | A kind of three-dimension gesture critical point detection method and system neural network based | |
CN110288695B (en) | Single-frame image three-dimensional model surface reconstruction method based on deep learning | |
CN108921926B (en) | End-to-end three-dimensional face reconstruction method based on single image | |
US6795069B2 (en) | Free-form modeling of objects with variational implicit surfaces | |
CN100407798C (en) | Three-dimensional geometric mode building system and method | |
CN110458939A (en) | The indoor scene modeling method generated based on visual angle | |
Ding et al. | A survey of sketch based modeling systems | |
CN111583408B (en) | Human body three-dimensional modeling system based on hand-drawn sketch | |
CN110633628B (en) | RGB image scene three-dimensional model reconstruction method based on artificial neural network | |
CN110399809A (en) | The face critical point detection method and device of multiple features fusion | |
Jin et al. | 3d reconstruction using deep learning: a survey | |
CN110349247A (en) | A kind of indoor scene CAD 3D method for reconstructing based on semantic understanding | |
Clarke et al. | Automatic generation of 3D caricatures based on artistic deformation styles | |
CN113822993A (en) | Digital twinning method and system based on 3D model matching | |
CN113345106A (en) | Three-dimensional point cloud analysis method and system based on multi-scale multi-level converter | |
Afifi et al. | Pixel2Point: 3D object reconstruction from a single image using CNN and initial sphere | |
CN111460193A (en) | Three-dimensional model classification method based on multi-mode information fusion | |
CN113160275A (en) | Automatic target tracking and track calculating method based on multiple videos | |
CN114612938A (en) | Dynamic gesture recognition method based on multi-view three-dimensional skeleton information fusion | |
CN114299339A (en) | Three-dimensional point cloud model classification method and system based on regional correlation modeling | |
CN117274388A (en) | Unsupervised three-dimensional visual positioning method and system based on visual text relation alignment | |
Pohle-Fröhlich et al. | Roof Segmentation based on Deep Neural Networks. |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |