CN107507126A - A kind of method that 3D scenes are reduced using RGB image - Google Patents
A kind of method that 3D scenes are reduced using RGB image Download PDFInfo
- Publication number
- CN107507126A CN107507126A CN201710621981.5A CN201710621981A CN107507126A CN 107507126 A CN107507126 A CN 107507126A CN 201710621981 A CN201710621981 A CN 201710621981A CN 107507126 A CN107507126 A CN 107507126A
- Authority
- CN
- China
- Prior art keywords
- scenes
- rgb image
- reduced
- window
- image according
- 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
- 238000000034 method Methods 0.000 title claims abstract description 27
- 238000001514 detection method Methods 0.000 claims description 22
- 238000004422 calculation algorithm Methods 0.000 claims description 14
- 239000011159 matrix material Substances 0.000 claims description 9
- 230000011218 segmentation Effects 0.000 claims description 9
- 230000007704 transition Effects 0.000 claims description 9
- 238000013136 deep learning model Methods 0.000 claims description 6
- 238000012549 training Methods 0.000 claims description 6
- 238000000605 extraction Methods 0.000 claims description 4
- 238000005457 optimization Methods 0.000 claims description 4
- 238000005070 sampling Methods 0.000 claims description 3
- 238000013519 translation Methods 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 claims description 2
- 238000009826 distribution Methods 0.000 claims description 2
- 238000007689 inspection Methods 0.000 claims 1
- 238000013527 convolutional neural network Methods 0.000 description 6
- 230000001815 facial effect Effects 0.000 description 4
- 210000004556 brain Anatomy 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 3
- 235000013399 edible fruits Nutrition 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 239000010410 layer Substances 0.000 description 2
- 230000004888 barrier function Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 239000011248 coating agent Substances 0.000 description 1
- 238000000576 coating method Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 239000011229 interlayer Substances 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000009877 rendering Methods 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/08—Projecting images onto non-planar surfaces, e.g. geodetic screens
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to 3-dimensional structure prediction field, the method for more particularly to a kind of three-dimensional 3-dimensional structure using RGB image.This method mainly includes image instance segmenting device, 3D model generating means, and object relative position prediction meanss.The input data of this method is 2 dimension RGB images.Output data is the threedimensional model of whole environment.This method can help robot to be better understood from local environment, reduce robot movement, keep in obscurity and the difficulty of path planning.
Description
Technical field
Present invention design is related to 3-dimensional structure prediction field, more particularly to a kind of side that 3D scenes are reduced using RGB image
Method.
Background technology
Human brain has the ability of completion environment 3D information.See the one side of a 3D object, people can guess out other
Appearance simultaneously, and the 3D structures of whole object are generated in the brain.It is this to speculate complete 3D structures using partial information
Ability is the Context aware of the mankind, mobile, avoiding barrier, the very big help of offer such as path planning.Machine does not possess robust still
3D structure completion abilities.With the help of infrared range-measurement system etc., machine also can only obtain the 3D information of visible part.Mesh
It is preceding help machine realize the method for 3D structure completions be also concentrated mainly on using the 3D object models pre-defined in CAD storehouses come
Synthesize or be spliced into the 3D structures to match with object in environment.Such way makes the 3D structures of generation excessively single, not
General it can use in scene changeable in real life.The reaction of same pictures is seen with the mankind as shown in figure 3, testing a machine,
Figure is the 2D images of certain scene；Understanding of the machine shown in Fig. 4 in the case where there is depth information to image, the part being blocked
Information has no way of learning；Understanding of the mankind shown in Fig. 5 to image, the object that human brain can be blocked with completion, reduce 3D structures.Closely
Remained little over year with what the low branch fruit of supervised learning was plucked, unsupervised learning becomes study hotspot.GAN
(Generative Adversarial Network, generation confrontation network) and VAE (Variational Auto-Encoder,
Variation self-encoding encoder) etc. self-supervisory generation model it is of increased attention.Generation model can utilize latent space
Random vector, generate significant image or animation.Fig. 7-10 is illustrated the 2D facial images completion realized using GAN and imitated
Fruit, Fig. 7 and Fig. 9 are that input picture is the facial image with noise, and Fig. 8 and Figure 10 are that output image is after removing noise
Face.
Same generation model is also extended into the task of generation 3D structures, the 3D-VAE- proposed by MIT in 2016
GAN can utilize the RGB image of object to be inferred to the 3D rendering of object.As people, present machine, which also has, utilizes RGB
Image infers the ability of object 3D structures.But 3D-VAE-GAN can only predict the 3D structures of single body, for containing complexity
Environment, the scene blocked comprising object or helpless.
In summary, the 2D images of the scene comprising multiple objects how are utilized, the tomograph of scene is reduced, is one
Individual urgent problem to be solved.
The content of the invention
It is an object of the invention to provide a kind of method that 3D scenes are reduced using RGB image, to help machine to understand more things
Body, there is the complex environment blocked.
To achieve these goals, the present invention provides following technical scheme：A kind of side that 3D scenes are reduced using RGB image
Method, idiographic flow are：
Step S1:Read in RGB image；
Step S2:Object detection is carried out in image range, the object detected is demarcated using rectangular window；
Step S3:The image detected carries out example segmentation, obtains the masking-out of single body；
Step S4:The 3D shape of single body is predicted using deep learning model, input data is single body
Masking-out；
Step S5:Regression forecasting is carried out the relativeness each pair youngster's object in image；
Step S6:Relativeness structure figure between the object obtained using prediction；
Step S7:Global figure optimization is carried out, obtains optimal object 3-dimensional space arrangement mode；
Step S8:Obtain three-dimensional scenic.
Further, step S2 is specifically included：Feature extraction, production detection are carried out on image using deep learning model
Object candidate region, the classification of detection object window and the optimal presumption of the window's position.Its advantage is to utilize deep learning model
Obtained testing result precision is high, and error is small.
Further step S3 is specifically included：Feature up-sampling, profit are carried out inside object candidate region caused by step S2
The classification of the detection object window interior pixel scale obtained with the method for bilinearity difference, by classification identical pixel in window
Integration is the masking-out of the detection object.
Specifically, the algorithm in window used in the classification of progress pixel scale described in S3 steps is SVM.Its is beneficial
Effect is split on the basis of detection, greatly reduce it is multi-class between interfere with each other, improve pixel classifications
Quality.
Further step S4 is specifically included：The object masking-out obtained by step S3 is inputted, utilizes variation self-encoding encoder
(VAE) object masking-out is projected into latent variable space, latent variable is generated to corresponding 3-dimensional knot using generation confrontation network
Structure.
It has been related to a kind of regression algorithm for predicting the relativeness between object in further step S5.Its feature exists
In：The translation between two articles and rotation relationship are represented using 3D transition matrixes；Instructed using deep learning algorithm and big data
Practice to make the distribution of relativeness between the real object of Model approximation.
Figure in further step S6 is non-directed graph G (V, E), and wherein V represents summit, represents the collection of detection object herein
Close, E is the set on side, represents the transition matrix between detection object herein.The mode for obtaining G is to be speculated respectively using step S5
Transition matrix between two objects.
Further step S7 purpose is to scheme caused by Optimization Steps S6.
The present invention utilizes object detecting method, detects the multiple objects in image；Pixel scale is carried out to multiple objects
Segmentation, obtain the masking-out of multiple objects；On this basis, the relative position each two object is predicted, generates nothing
Xiang Tu；More objects puts relation in finally being optimized the environment using figure optimized algorithm；The present invention solves direct by two dimensional image
Infer the problem of three-D space structure.
Brief description of the drawings
Fig. 1 is that 3D of the present invention estimates algorithm flow chart.
Fig. 2 is the example segmentation effect figure based on MaskRCNN.
Fig. 3 is to convert images into 3D model structures using self-encoding encoder+maker.
Fig. 4 be test a machine the reaction that same pictures are seen with the mankind certain scene 2D images.
Fig. 5 is the figure of shown understanding of the machine in the case where there is depth information to image.
The figure of understanding of the mankind shown in Fig. 6 to image.
Fig. 7 be using GAN realize 2D input pictures be the facial image with noise.
The output image that Fig. 8 is Fig. 7 removes the face figure after noise.
Fig. 9 is that another 2D input pictures realized using GAN are the facial image with noise.
The output image that Figure 10 is Fig. 9 removes the face figure after noise.
Embodiment
The specific algorithm of each step of the invention is described below.
A kind of method that 3D scenes are reduced using RGB image, shown in idiographic flow Fig. 1；
The embodiment that object detection is carried out in image range involved by step S2 is as follows：
Object detection is carried out in image range, the object detected is demarcated using rectangular window.Here with
Detection mode based on Faster R-CNN algorithms.Concretely comprise the following steps：Feature extraction, produce candidate region, window classification and essence
Repair.Characteristic extraction part is realized using depth convolutional neural networks (CNN).If CNN include comprising dried layer volume basic unit (conv) and
Active coating (relu).The VGG networks that the CNN of this programme trains to obtain using ImageNet.Faster RCNN utilize RPN (candidates
Area generation network) generation candidate region.In Faster RCNN, PRN be a very little convolutional network (3 × 3conv →
The conv of 1 × 1conv → 1 × 1) all characteristic quantities of conv5_3 volumes of basic unit's characteristic pattern are checked in moving window.It is each mobile
Window has 9 with the related anchor box (anchor box) of its receptive field.PRN can do the window's position to each anchor box
Return (bounding box regression) and window confidence score (box confidence scoring).Pass through knot
Closing the loss of above three turns into common all characteristic quantity, and whole pipeline can be trained to.Faster R-CNN top
Layer output also includes bounding box regression and window confidence score, and the two parts are respectively intended to refine window
Mouthful, and the object in window is identified/classified.
The image progress example segmentation to detecting involved by step S3, obtains the specific implementation of the masking-out of single body
Mode is as follows：
Example refers to 2D images being subdivided into multiple images subregion (set of pixel), wherein per sub-regions
Pixel belong to same object.Scheme utilizes the Mask-RCNN of Facebook AI research institutes.Fig. 2 is the one of example segmentation
Individual effect.For each object in image, example segmentation can generate a masking-out (Mask)., can be by image using masking-out
In object segmentation come out.Mask RCNN are improved based on faster RCNN.By increasing a particular category pair
As mask is predicted, Mask RCNN extend the Faster RCNN of facing living examples segmentation, with existing bounding box regressor and right
As grader is parallel.Because Faster RCNN RolPool is not designed for the pixel of network inputs and outlet chamber to picture
Element alignment, MaskRCNN instead of it with RolAlign.RolAlign has used bilinear interpolation to calculate each subwindow
The exact value of input feature vector, rather than RolPooling maximum pond method.MaskRCNN training dataset is storehouse of increasing income
Microsoft COCO。
The embodiment that the 3D shape of single body is predicted using deep learning model that step S4 is related to
It is as follows：
The 3D shape of single body is predicted using VAE-GAN, input data is the masking-out of single body.VAE (variations
Self-encoding encoder) with come on, image is mapped to latent variable space (latent space) and then again by the data in latent variable space
It is mapped to a kind of mode of own coding decoding of image.GAN (generation confrontation network) is given birth to using the data in latent variable space
Into the method for True Data.VAE own coding part is grafted onto before GAN by this programme, it is therefore an objective to is reflected 2D images using VAE
Latent variable space is mapped to, then latent variable is mapped to above 3D models using GAN.VAE-GAN includes three parts：Figure
As encoder：E, decoder：GAN, and an identifier：D.Image encoder include five spaces roll up machine kernels be respectively 11,
5,5,5,8 } and step-length { 4,2,2,2,1 } convolution interlayer is using crowd standardization (batch normalization) and ReLU, and
One sampler is used for sampling the vector of 200 dimensions.VAE-GAN structure as shown in figure 3, input example masking-out → self-encoding encoder →
Latent variable → maker → reconstruct 3D shapes.
The loss function L that VAE-GAN learning process is related to includes three parts：Object reconstruction loses Lrecon, intersects
Entropy loss L3D-GAN and relative entropy loss LKL.It is written as equation：
L=L3D-GAN+a1LKL+a2Lrecon,
The weight that wherein a1 and a2 is reconstruct loss respectively and relative damage entropy loses.We set a1=5, a2=104.
The calculation equation of wherein three losses is as follows：
L3D-GAN=log D (x)+log (1D (G (z))),
LKL=DKL (q (z | y) | | p (z)),
Lrecon=| | G (E (y)) x | | 2,
Wherein x is the 3D models that training data is concentrated, and y is 2D models on the other side, and q (z | y) it is latent space z
Variable probability is distributed.This programme selection p (z) is distributed for a multivariate Gaussian, and its average is 0, and variance is 1. training VAE-GAN
When need 2D images and corresponding 3D models while exist.The training dataset is utilized to the manual demarcation of CAD model
Obtain.
The specific embodiment party that regression forecasting is carried out the relativeness each pair youngster's object in image that step S5 is related to
Formula is as follows：
Regression forecasting is carried out the relativeness each pair youngster's object in image.Obtain each object 3D shapes it
Afterwards.Finally utilize the relativeness returned between device prediction 3D models.The conversion of relativeness 3d space between 3D models
Matrix T is represented.T is 4x4 matrix, and the 3x3 submatrixs in the wherein T upper left corner represent the rotation pass of three dimensions, T
The 4th row [Xt, Yt, Zt] represent three dimensions in translation relation.The target of regression algorithm is exactly to minimize transition matrix
Difference between predicted value and actual value.The present invention represents this error using L2 losses.The method for expressing of L2 losses is true
The square value of difference between value and predicted value.Regression algorithm profit MLP (multilayer neural network).Transition matrix in training data
Actual value is directly to calculate to obtain in the 3D models of simulation.
The embodiment for the figure optimization being related in step S7 is as follows.
The transition matrix of detection object two-by-two is calculated in step S6, and has obtained non-directed graph G (V, E), wherein V
Summit is represented, represents the set of detection object herein, E is the set on side, represents the transformational relation between detection object herein.Such as
Occurs two or more object in fruit environment, due to all including the relative position relation of supposition between two two articles, so by difference
Detection object recursion is to can produce control information between the putting position of all objects.Be described below how selectivity use nothing
To figure G side collection to eliminate this error.
Deep neural network has been used in step S2 to do object detection algorithms.For each object detected, god
A confidence level C can be exported through network：Confidence, to represent the degree of reliability of this object being detected.We are sharp
The side collection of non-directed graph is screened with this confidence level：
Three dimensions puts algorithm
Claims (9)
- A kind of 1. method that 3D scenes are reduced using RGB image, it is characterised in that：Concretely comprise the following steps：Step S1:Read in RGB image；Step S2:Object detection is carried out in image range, the object detected is demarcated using rectangular window；Step S3:The image detected carries out example segmentation, obtains the masking-out of single body；Step S4:The 3D shape of single body is predicted using deep learning model, input data is the masking-out of single body；Step S5:Regression forecasting is carried out the relativeness each pair youngster's object in image；Step S6:Relativeness structure figure between the object obtained using prediction；Step S7:Global figure optimization is carried out, obtains optimal object 3-dimensional space arrangement mode；Step S8:Obtain three-dimensional scenic.
- A kind of 2. method that 3D scenes are reduced using RGB image according to claim 1, it is characterised in that：The step S2 is specifically included：Feature extraction is carried out on image using deep learning model, produces detection object candidate Region, the classification of detection object window and the optimal presumption of the window's position.
- A kind of 3. method that 3D scenes are reduced using RGB image according to claim 1, it is characterised in that：The step S3 is specifically included：Feature up-sampling is being carried out caused by the step S2 inside object candidate region, is being utilized The classification for the detection object window interior pixel scale that the method for bilinearity difference obtains, classification identical pixel in window is united It is combined into the masking-out of the detection object.
- A kind of 4. method that 3D scenes are reduced using RGB image according to claim 3, it is characterised in that：S3 steps institute The algorithm used in the classification that pixel scale is carried out in window stated is SVM.
- A kind of 5. method that 3D scenes are reduced using RGB image according to claim 1, it is characterised in that：The step S4 is specifically included：The object masking-out that is obtained by the step S3 is inputted, using variation self-encoding encoder by object Masking-out projects to latent variable space, and latent variable is generated to corresponding 3-dimensional structure using generation confrontation network.
- A kind of 6. method that 3D scenes are reduced using RGB image according to claim 1, it is characterised in that：It is related to a kind of regression algorithm for predicting the relativeness between object in the step S5, is represented using 3D transition matrixes Translation and rotation relationship between two articles；Make the real thing of Model approximation using deep learning algorithm and big data training The distribution of relativeness between body.
- A kind of 7. method that 3D scenes are reduced using RGB image according to claim 1, it is characterised in that：The step Figure in S6 is G (V, E), and wherein V represents summit, represents the set of detection object herein, and E is the set on side, represents inspection herein The transition matrix surveyed between object.
- A kind of 8. method that 3D scenes are reduced using RGB image according to claim 7, it is characterised in that：Obtain G side Formula is to speculate the transition matrix between two objects respectively using the step S5.
- A kind of 9. method that 3D scenes are reduced using RGB image according to claim 1, it is characterised in that：The step S7 purpose is schemed to optimize caused by the step S6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710621981.5A CN107507126B (en) | 2017-07-27 | 2017-07-27 | Method for restoring 3D scene by using RGB image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710621981.5A CN107507126B (en) | 2017-07-27 | 2017-07-27 | Method for restoring 3D scene by using RGB image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107507126A true CN107507126A (en) | 2017-12-22 |
| CN107507126B CN107507126B (en) | 2020-09-18 |
Family
ID=60689937
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710621981.5A Active CN107507126B (en) | 2017-07-27 | 2017-07-27 | Method for restoring 3D scene by using RGB image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107507126B (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108304831A (en) * | 2018-03-15 | 2018-07-20 | 广东工业大学 | A kind of method and device that monitoring worker safety helmet is worn |
| CN108388880A (en) * | 2018-03-15 | 2018-08-10 | 广东工业大学 | A kind of method and device that monitoring driver drives using mobile phone |
| CN108495110A (en) * | 2018-01-19 | 2018-09-04 | 天津大学 | A kind of virtual visual point image generating method fighting network based on production |
| CN108648197A (en) * | 2018-04-12 | 2018-10-12 | 天津大学 | A kind of object candidate area extracting method based on image background mask |
| CN108875818A (en) * | 2018-06-06 | 2018-11-23 | 西安交通大学 | Based on variation from code machine and confrontation network integration zero sample image classification method |
| CN108932693A (en) * | 2018-06-15 | 2018-12-04 | 中国科学院自动化研究所 | Face editor complementing method and device based on face geological information |
| CN108959551A (en) * | 2018-06-29 | 2018-12-07 | 北京百度网讯科技有限公司 | Method for digging, device, storage medium and the terminal device of neighbour's semanteme |
| CN109145769A (en) * | 2018-08-01 | 2019-01-04 | 辽宁工业大学 | The target detection network design method of blending image segmentation feature |
| CN109191414A (en) * | 2018-08-21 | 2019-01-11 | 北京旷视科技有限公司 | A kind of image processing method, device, electronic equipment and storage medium |
| CN109472795A (en) * | 2018-10-29 | 2019-03-15 | 三星电子（中国）研发中心 | A kind of image edit method and device |
| CN110648305A (en) * | 2018-06-08 | 2020-01-03 | 财团法人工业技术研究院 | Industrial image detection method, system and computer readable recording medium |
| CN110706328A (en) * | 2019-08-21 | 2020-01-17 | 重庆特斯联智慧科技股份有限公司 | Three-dimensional scene virtual generation method and system based on GAN network |
| CN112505049A (en) * | 2020-10-14 | 2021-03-16 | 上海互觉科技有限公司 | Mask inhibition-based method and system for detecting surface defects of precision components |
| CN112712460A (en) * | 2020-12-09 | 2021-04-27 | 杭州妙绘科技有限公司 | Portrait generation method and device, electronic equipment and medium |
| WO2022240354A1 (en) * | 2021-05-14 | 2022-11-17 | Lemon Inc. | A high-resolution portrait stylization frameworks using a hierarchical variational encoder |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101714262A (en) * | 2009-12-10 | 2010-05-26 | 北京大学 | Method for reconstructing three-dimensional scene of single image |
| CN105930382A (en) * | 2016-04-14 | 2016-09-07 | 严进龙 | Method for searching for 3D model with 2D pictures |
| CN106709568A (en) * | 2016-12-16 | 2017-05-24 | 北京工业大学 | RGB-D image object detection and semantic segmentation method based on deep convolution network |
-
2017
- 2017-07-27 CN CN201710621981.5A patent/CN107507126B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101714262A (en) * | 2009-12-10 | 2010-05-26 | 北京大学 | Method for reconstructing three-dimensional scene of single image |
| CN105930382A (en) * | 2016-04-14 | 2016-09-07 | 严进龙 | Method for searching for 3D model with 2D pictures |
| CN106709568A (en) * | 2016-12-16 | 2017-05-24 | 北京工业大学 | RGB-D image object detection and semantic segmentation method based on deep convolution network |
Non-Patent Citations (1)
| Title |
|---|
| METAL1: "Facebook的图像识别很强大，一次开源了三款机器视觉工具", 《HTTPS://BLOG.CSDN.NET/METAL1/ARTICLE/DETAILS/53318437》 * |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108495110B (en) * | 2018-01-19 | 2020-03-17 | 天津大学 | Virtual viewpoint image generation method based on generation type countermeasure network |
| CN108495110A (en) * | 2018-01-19 | 2018-09-04 | 天津大学 | A kind of virtual visual point image generating method fighting network based on production |
| CN108388880A (en) * | 2018-03-15 | 2018-08-10 | 广东工业大学 | A kind of method and device that monitoring driver drives using mobile phone |
| CN108304831B (en) * | 2018-03-15 | 2022-03-22 | 广东工业大学 | Method and device for monitoring wearing of safety helmet of worker |
| CN108304831A (en) * | 2018-03-15 | 2018-07-20 | 广东工业大学 | A kind of method and device that monitoring worker safety helmet is worn |
| CN108648197B (en) * | 2018-04-12 | 2021-09-07 | 天津大学 | Target candidate region extraction method based on image background mask |
| CN108648197A (en) * | 2018-04-12 | 2018-10-12 | 天津大学 | A kind of object candidate area extracting method based on image background mask |
| CN108875818A (en) * | 2018-06-06 | 2018-11-23 | 西安交通大学 | Based on variation from code machine and confrontation network integration zero sample image classification method |
| CN110648305A (en) * | 2018-06-08 | 2020-01-03 | 财团法人工业技术研究院 | Industrial image detection method, system and computer readable recording medium |
| US11315231B2 (en) | 2018-06-08 | 2022-04-26 | Industrial Technology Research Institute | Industrial image inspection method and system and computer readable recording medium |
| CN108932693A (en) * | 2018-06-15 | 2018-12-04 | 中国科学院自动化研究所 | Face editor complementing method and device based on face geological information |
| CN108932693B (en) * | 2018-06-15 | 2020-09-22 | 中国科学院自动化研究所 | Face editing and completing method and device based on face geometric information |
| CN108959551B (en) * | 2018-06-29 | 2021-07-13 | 北京百度网讯科技有限公司 | Neighbor semantic mining method and device, storage medium and terminal equipment |
| CN108959551A (en) * | 2018-06-29 | 2018-12-07 | 北京百度网讯科技有限公司 | Method for digging, device, storage medium and the terminal device of neighbour's semanteme |
| CN109145769A (en) * | 2018-08-01 | 2019-01-04 | 辽宁工业大学 | The target detection network design method of blending image segmentation feature |
| CN109191414A (en) * | 2018-08-21 | 2019-01-11 | 北京旷视科技有限公司 | A kind of image processing method, device, electronic equipment and storage medium |
| CN109472795A (en) * | 2018-10-29 | 2019-03-15 | 三星电子（中国）研发中心 | A kind of image edit method and device |
| CN110706328A (en) * | 2019-08-21 | 2020-01-17 | 重庆特斯联智慧科技股份有限公司 | Three-dimensional scene virtual generation method and system based on GAN network |
| CN112505049B (en) * | 2020-10-14 | 2021-08-03 | 上海互觉科技有限公司 | Mask inhibition-based method and system for detecting surface defects of precision components |
| CN112505049A (en) * | 2020-10-14 | 2021-03-16 | 上海互觉科技有限公司 | Mask inhibition-based method and system for detecting surface defects of precision components |
| CN112712460A (en) * | 2020-12-09 | 2021-04-27 | 杭州妙绘科技有限公司 | Portrait generation method and device, electronic equipment and medium |
| CN112712460B (en) * | 2020-12-09 | 2024-05-24 | 杭州妙绘科技有限公司 | Portrait generation method, device, electronic equipment and medium |
| WO2022240354A1 (en) * | 2021-05-14 | 2022-11-17 | Lemon Inc. | A high-resolution portrait stylization frameworks using a hierarchical variational encoder |
| US11720994B2 (en) | 2021-05-14 | 2023-08-08 | Lemon Inc. | High-resolution portrait stylization frameworks using a hierarchical variational encoder |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107507126B (en) | 2020-09-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107507126A (en) | A kind of method that 3D scenes are reduced using RGB image | |
| Bozic et al. | Transformerfusion: Monocular rgb scene reconstruction using transformers | |
| Ge et al. | 3d hand shape and pose estimation from a single rgb image | |
| KR102610030B1 (en) | Deep learning system for cuboid detection | |
| CN107818580A (en) | 3D reconstructions are carried out to real object according to depth map | |
| Saputra et al. | Learning monocular visual odometry through geometry-aware curriculum learning | |
| CN108121995A (en) | For identifying the method and apparatus of object | |
| Ranjan et al. | Learning human optical flow | |
| EP4377898A1 (en) | Neural radiance field generative modeling of object classes from single two-dimensional views | |
| Nozawa et al. | 3D car shape reconstruction from a contour sketch using GAN and lazy learning | |
| US20230419600A1 (en) | Volumetric performance capture with neural rendering | |
| Peng | Machines' perception of space | |
| US11138812B1 (en) | Image processing for updating a model of an environment | |
| Rabby et al. | BeyondPixels: A comprehensive review of the evolution of neural radiance fields | |
| Yu et al. | NF-atlas: Multi-volume neural feature fields for large scale lidar mapping | |
| CN116134491A (en) | Multi-view neuro-human prediction using implicit differentiable renderers for facial expression, body posture morphology, and clothing performance capture | |
| JP2023079022A (en) | Information processing device and information generation method | |
| CN116385667B (en) | Reconstruction method of three-dimensional model, training method and device of texture reconstruction model | |
| Correia et al. | 3D reconstruction of human bodies from single-view and multi-view images: A systematic review | |
| CN116758212A (en) | 3D reconstruction method, device, equipment and medium based on self-adaptive denoising algorithm | |
| CN116434343A (en) | Video motion recognition method based on high-low frequency double branches | |
| CN116883524A (en) | Image generation model training, image generation method and device and computer equipment | |
| Williams | Structured Generative Models for Scene Understanding | |
| Van de Maele et al. | Object-centric scene representations using active inference | |
| Zeng et al. | 3D Reconstruction of buildings based on transformer-MVSNet |
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 | ||
| CB02 | Change of applicant information |
Address after: 116023 Room 1 203, Block 12, Block C, 6 Yiyang Road, Qixianling, Dalian High-tech Industrial Park, Liaoning Province Applicant after: Hechuang Lazy Man (Dalian) Technology Co.,Ltd. Address before: 116023 Dalian High-tech Park, Liaoning Province, Takeoff Software Park Phase II 1-04B Applicant before: DALIAN HECHUANG LANREN TECHNOLOGY CO.,LTD. |
|
| CB02 | Change of applicant information | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20220328 Address after: 116000 unit 07, 16 / F, animation tour industry building, No. 10, Torch Road, Dalian high tech Industrial Park, Dalian, Liaoning Province Patentee after: Dalian Yongge Technology Co.,Ltd. Address before: 116023 Room 1 203, Block 12, Block C, 6 Yiyang Road, Qixianling, Dalian High-tech Industrial Park, Liaoning Province Patentee before: Hechuang Lazy Man (Dalian) Technology Co.,Ltd. |
|
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20240108 Address after: No. 532-5, 5th Floor, World Trade Building, No. 25 Tongxing Street, Zhongshan District, Dalian City, Liaoning Province, 116001 Patentee after: Yuexin (Dalian) Technology Co.,Ltd. Address before: 116000 unit 07, 16 / F, animation tour industry building, No. 10, Torch Road, Dalian high tech Industrial Park, Dalian, Liaoning Province Patentee before: Dalian Yongge Technology Co.,Ltd. |
|
| TR01 | Transfer of patent right |