CN108876907A - A kind of active three-dimensional rebuilding method of object-oriented object - Google Patents
A kind of active three-dimensional rebuilding method of object-oriented object Download PDFInfo
- Publication number
- CN108876907A CN108876907A CN201810576919.3A CN201810576919A CN108876907A CN 108876907 A CN108876907 A CN 108876907A CN 201810576919 A CN201810576919 A CN 201810576919A CN 108876907 A CN108876907 A CN 108876907A
- Authority
- CN
- China
- Prior art keywords
- visual angle
- module
- network
- dimensional
- target
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
-
- 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)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Graphics (AREA)
- Geometry (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention belongs to technical field of computer vision, provide a kind of active three-dimensional rebuilding method of object-oriented object, specifically include following two module:Visual angle dynamic prediction module and Target self-determination rebuild module;Two modules all carry out following steps:Module input, module architectures and training method.The present invention is in order to solve conventional three-dimensional Object reconstruction vulnerable to environmental factor interference, inefficiency and the technical problem for being difficult to realize independence, devise the autonomous reconstruction framework of an objective based on depth learning technology and software platform, it can be to given target, Dynamic Programming scans visual angle, and in combination with the picture under different perspectives, the building of target three-dimensional is completed.
Description
Technical field
The invention belongs to technical field of computer vision, more particularly to are carried out independently based on deep learning to single target
The method of three-dimensional reconstruction.
Background technique
With the development of SLAM technology (Simultaneous Localization and Mapping), indoor scene
Three-dimensional rebuilding method reaches its maturity.Three-dimensional reconstruction generally comprises three parts, is carried out first using handheld camera to target to be reconstructed
Then the scanning at multiple visual angles carries out the extraction, matching and the estimation of camera pose of feature, finally to the multiframe picture scanned
Mapping of the two-dimensional pixel to three-dimensional coordinate point, the model finally rebuild are completed by stereovision technique.However, previous
Work in, to the scanning of target often using the scanning at a kind of " no dead angle ", i.e. scanning each office for wanting coverage goal
Portion's structure, inefficiency while, which also results in, can not be suitable for other types to the planning parameters of scanning paths of a certain target, to sweep
The independence retouched brings challenges.Therefore, a kind of the efficient reconstruction method at contexture by self scanning visual angle to be in reconstruction process
The invention motivation of technical barrier and this patent currently to be captured.Next relevant background in this field is discussed in detail
Technology.
(1) three-dimensional reconstruction
Early in 2010, University of Washington and Microsoft laboratory were developed based on SIFT (Scale invariant features transform) spy
Sign matching positioning and the real-time vision SLAM (positioning immediately of TORO (Tree-based Network Optimizer) optimization algorithm
With map structuring) system by this real-time system can establish out the three-dimensional map of scene.Then there are many work in reality
It is improved on Shi Xing, pose estimation, including RGBD-SLAM algorithm, KinectFusion, BundleFusion etc., it is substantially full
The real-time that foot scene rebuilding is interacted with user.
However these algorithms will inevitably face following problem:First, algorithm needs dense visual angle, and
It is practical calculate in but skip most of visual angle so that the acquisition inefficiency of information, and can not apply and having the more field blocked
Jing Zhong;Second, algorithm needs to assume that target does not have mirror-reflection in scene, while possessing texture abundant, to meet to picture
Feature extraction;Third, although having used a variety of optimisation strategies, the accumulation problem of camera registration error is still remained.
To solve the above-mentioned problems, Choy et al. works the three-dimensional reconstruction that deep learning introduces target, proposes three
The Recognition with Recurrent Neural Network of dimension, by building three-dimensional hiding layer state, to receive the picture shot under multiple visual angles, and meanwhile it is recessive
Indicate the geometry currently rebuild, to establish single-view and the united three-dimensional reconstruction frame of multi-angle of view, single-view with
Multi-angle of view (less than 20 visual angles) can show the effect of beyond tradition method.
The application of depth learning technology is that three-dimensional reconstruction opens new approaches, compares conventional method, uses less vision
Input, is capable of handling complex environment factor, and the combination of conventional method and deep learning thought may be the following three-dimensional reconstruction neck
The new guiding of one of domain.
(2) the autonomous acquisition of three-dimensional information
Robot below one arbitrarily visual angle to unknown target, it is next come active predicting based on current observation
Observation visual angle, and visual angle is estimated based on the observation at next visual angle, such a continuous view prediction can help machine
The interested information of people's actively perceive, completes corresponding visual task.Active sense is carried out to ambient enviroment by consumer level camera
Know, and the environmental information digitlization that will acquire is technical barrier to be captured in robot field.In previous work, often
Make to receive information maximization by the constraint in some rules.Common method includes that entropy is reduced, and uncertainty is reduced, Meng Teka
Sieve sampling, Gaussian process recurrence etc..
Also there is using the method for deep learning the selection for predicting visual angle in recent years, such as network-evaluated using depth confidence
The information gain at visual angle, visual attention model based on enhancing study etc..Its middle finger it is emphasised that the three-dimensional of Xu Kai et al.
Target self-determination identifies work.The worked combination multi-angle of view convolutional neural networks [24] and circulation attention model, realize mesh
Mark does not in the process obtain the active of depth data, and introduces space in subsequent work and migrate network, realizes end and arrives
The study at end.The actively perceive of the program be embodied in for current visual angle obtain visual observation, can predict it is next most
Good visual angle realizes the view prediction of view-based access control model feedback, allows the robot to the identification for independently completing target.
In view of the foregoing it is apparent that current generation, autonomous this problem of the visual information also still place's exploration perceived in environment
Stage, current work are mainly placed on center of gravity in the identification and retrieval of target, also need in more complicated Object reconstruction problem
It further to explore.
Summary of the invention
The present invention in order to solve conventional three-dimensional Object reconstruction vulnerable to environmental factor interference, inefficiency and be difficult to realize from
The technical problem of main property devises the autonomous reconstruction framework of an objective based on depth learning technology and software platform, energy
Enough to given target, Dynamic Programming scans visual angle, and in combination with the picture under different perspectives, completes target three-dimensional
Building.
Technical solution of the present invention:
A kind of active three-dimensional rebuilding method of object-oriented object, specifically includes following two module:
(1) visual angle dynamic prediction module:
(1.1) module inputs:
Carry out target information in collection room using RGBD camera, any arbitrary viewing angles v around target0, with RGBD camera
Shooting obtains photochrome, using wherein tri- channels RGB and by resolution compression to 64 × 64, obtains 64 × 64 × 3
Measure I0, arbitrary viewing angles v0With picture tensor I0Collectively form the input of module;
(1.2) module architectures:
Visual angle dynamic prediction module is the neural network of a Ge Liang branch, in different times walk under have different states,
It outputs and inputs;It is sometime walking in t, the first branching networks are a full articulamentum fview, with arbitrary viewing angles vtAs defeated
Enter, calculates corresponding visual angle characteristic fview(v0);Second branching networks are a multilayer convolution loop neural network fenc, become
Coding network is responsible for that picture tensor I will be inputtedtIt is encoded to the feature of low-dimensionalWhereinIt is circulation
Layer is in the storage state of a upper time step, and in t=1, value is unit matrix;
The feature that then two branching networks are extracted carries out the multiplication of Element-Level, and in another circulation layer fgruWith it is complete
Articulamentum ffcProcessing under, obtain final feature vector such as formula (1):
Wherein,It is storage state of the circulation layer in a upper time step, in t=1, value is unit matrix;
Finally, feature vector F to be passed through to the processing of S function (sigmoid), the visual angle finally predicted, as next
The input viewing angle v of time stept+1, while the picture under the visual angle is obtained using RGBD camera, obtain the input of future time step
Measure It+1;
(1.3) training method:
The training of neural network is pressed in each time step using enhancing learning method in the dynamic prediction module of visual angle
Illuminated (2) calculates reward:
Wherein,For the Three-dimension Reconstruction Model under t moment, V is the true model in database, and IoU is two model weights
Folded element number accounts for the specific gravity of whole elements;
After the reward for each time step that adds up, the optimization method based on strategy estimation, training network are used;It is specifically used
The gradient that the method for gradient decline calculates then updates network parameter in the direction iteration that gradient reduces, obtains pre- such as formula (3)
Survey the neural network at optimal visual angle:
Wherein, R is the cumulative award of all time steps,The probability at the visual angle is obtained for prediction;
(2) Target self-determination rebuilds module:
(2.1) module inputs:
Picture is shot under the visual angle predicted in the dynamic prediction module of visual angle using RGBD camera, it will according to time step sequence
Every frame picture obtains orderly picture tensor sequence { I for identical processing method in visual angle dynamic prediction module0, I1...,
In, the input as module;
(2.2) module architectures:
It is a Recognition with Recurrent Neural Network that Target self-determination, which rebuilds module, includes two parts of coding network and decoding network;It compiles
Code network fencUsing the coding network (second point to network) of visual angle dynamic prediction module, i.e. a multilayer convolution loop nerve
Network is responsible for the picture tensor I inputted under coding t time steptLow-dimensional feature Ft;Decoding network fdecBy the three-dimensional warp of multilayer
Product circulation layer composition, by the picture feature F of low-dimensionaltIt rises dimension and obtains three-dimensional voxelWherein
It is storage state of the circulation layer in a upper time step, in t=1, value is unit matrix;
The sequence of pictures being made of for one t time step is sequentially input to visual angle dynamic prediction mould according to time step
In block, the three-dimensional voxel predicted under the last one (i.e. t-th) time step is the three-dimensional reconstruction result of target;
(2.3) training method:
Target self-determination rebuild module using direction propagate with the training of the method for stochastic gradient descent, for a lot sample sheet,
The error between the prediction result of network and database true value result is calculated according to formula (4), and calculates the gradient of error, according to
The backpropagation of neural network gradually updates network parameter along the direction of gradient decline, and iteration is until convergence;
Lvox=| | Vpre-V||2 (4)
Wherein, VpreIt is respectively indicated with V corresponding in the voxel model and database of Target self-determination reconstruction module network prediction
Voxel model.
The present invention has outstanding feature compared with similar, and specific detailed description are as follows:
(1) towards the independence of reconstruction process
It is different from traditional scanning mode at " no dead angle ", independence of this patent towards reconstruction process.It can specifically show
At two aspects.First, other method for reconstructing mostly use pre-specified path, using sensor to each part of target
Structure is scanned, and the three-dimensional information of target is calculated by processing multiframe scanning information.Which results in a certain target
The route of design not can be used directly other class method for distinguishing generally, and universality is not strong, and especially blocking for environment may
Hinder the planning in some paths, it is difficult to realize planning from main perspective for reconstruction.And the dynamic visual angle prediction of this patent can be directed to
Different types of target carries out the prediction and Target Modeling at visual angle simultaneously in reconstruction process, realizes independence;Second, other
Method for reconstructing determines scanning the obtained information content in visual angle often through the measurement to reconstructed results, and this patent is using a kind of
The mode of " seeing i.e. gained ", without reconstruction, directly judges next optimal viewing angle, raising efficiency for the photo of shooting
While to assign entire automated processes more " intelligence ".
(2) trial of depth learning technology
From traditional multiframe match it is different with the method for stereoscopic vision, this patent trial use depth learning technology to target into
Row three-dimensional modeling, this results in the advantages of two aspects.First, other methods generally require to calculate the matching spy for closing on interframe
Sign, therefore it is required that number viewpoints guarantee that the overlapping region for closing on interframe is enough big more.And deep learning method due to itself
Data-driven ability, can be used the information under a certain visual field of target to predict the information under other visuals field and global structure,
The three-dimensional reconstruction for completing target using few visual angle is enabled the system to, the effect of system is improved while improving robustness
Rate;Second, other methods are limited by environmental factor, it is difficult to extract under intense light irradiation, mirror-reflection, obstruction conditions enough
Feature is unable to reach ideal reconstruction effect to increase the error of camera pose estimation.However depth learning technology itself
With the ability for predicting global information from local message, for rugged environment factor, the spy of robust often can be also extracted
Sign, obtains preferable reconstructed results.
Detailed description of the invention
Fig. 1 is concept database figure.Database includes the magnanimity target data of multiple classifications, and different target is by individual rope
It is incorporated in retrieval, and each target separately includes the data of two kinds of forms, is that (whole visual angles need to cover at multiple visual angles respectively
All structures of target) under the two-dimension picture rendered and voxelization three-dimensional voxel.
Fig. 2 is network architecture diagram.The network architecture diagram of autonomous reconstruction model has only drawn the data flow of a time step in figure
Emotionally condition, network include that Target self-determination rebuilds module and visual angle dynamic prediction module, are respectively used to receive reconstruction model and prediction
Next visual angle.
Fig. 3 is autonomous reconstruction flow chart.Scheming a indicates that visual angle camera position in Dynamic selection processes converts, and figure b indicates whole
A flow chart independently rebuild originates in an arbitrary viewing angles, shoots the picture under the visual angle, respectively by trained in advance
Visual angle dynamic prediction module and Target self-determination rebuild module.The former is responsible for receiving picture and current visual angle, judge one it is best
Visual angle;The latter is for rebuilding current threedimensional model.This is a duplicate process, and visual angle dynamic prediction module repeatedly judges
Next visual angle is to obtain the input of next picture, and Target self-determination rebuilds module then constantly reconstruction model, until iteration ends.
Specific embodiment
Invention is further described in detail With reference to embodiment, but the invention is not limited to specific realities
Apply mode.
A method of autonomous three-dimensional reconstruction being carried out to single target based on deep learning, the training including network model
And the autonomous reconstruction part of model
(1) training network model
A large-scale database is constructed first, is included different classes of magnanimity three-dimensional grid model in database, is passed through
Multi-angle of view rendering is carried out to grid model and grid voxelization obtains training data (the three-dimensional voxel model comprising target and more views
Angle picture illustrates visible Fig. 1), while network model is built according to Fig. 2;Then training data multithreading is transported in batches
In network model to be trained, and visual angle dynamic prediction module and Target self-determination reconstruction module are calculated according to formula (3) and formula (4)
In error;Finally decline optimizer iteration according to the method for backpropagation gradient and update network parameter, until the number of iterations
It meets the requirements, completes the training of network.
(2) active reconstruction process
The target to be reconstructed for one, operate machine people or human hand held RGBD camera come it is a certain random around target
Position, and guarantee that there are targets in the camera fields of view under the visual angle.It is shot with camera and obtains photo, it is logical using wherein RGB tri-
Road and the compression for carrying out resolution ratio are input in housebroken network, and at this moment Target self-determination, which rebuilds module, can export current reconstruction
Good model, and visual angle dynamic prediction module can export next optimal viewing angle.Robot or human hand held RGBD camera are come down
One optimal viewing angle continues to shoot picture, be transported in housebroken network, and obtains the model rebuild and next view
Angle.This process repeats, until the model rebuild meet the requirements or visual angle is defeated have reached preset threshold value when termination,
The structure that Target self-determination rebuilds module output at this time is final Object reconstruction model, illustrates visible Fig. 3.
Claims (1)
1. a kind of active three-dimensional rebuilding method of object-oriented object, which is characterized in that active three-dimensional rebuilding method has
Two modules:
(1) visual angle dynamic prediction module:
(1.1) module inputs:
Carry out target information in collection room using RGBD camera, any arbitrary viewing angles v around target0, shot with RGBD camera
64 × 64 × 3 tensor I is obtained using wherein tri- channels RGB and by resolution compression to 64 × 64 to photochrome0, with
Machine visual angle v0With picture tensor I0Collectively form the input of module;
(1.2) module architectures:
Visual angle dynamic prediction module is the neural network of a Ge Liang branch, there is different states, input under walking in different times
And output;It is sometime walking in t, the first branching networks are a full articulamentum fview, with arbitrary viewing angles vtAs input, meter
Calculate corresponding visual angle characteristic fview(v0);Second branching networks are a multilayer convolution loop neural network fenc, become coding
Network is responsible for that picture tensor I will be inputtedtIt is encoded to the feature of low-dimensionalWhereinIt is circulation layer
In the storage state of a upper time step, in t=1, value is unit matrix;
The feature that then two branching networks are extracted carries out the multiplication of Element-Level, and in another circulation layer fgruWith full connection
Layer ffcProcessing under, obtain final feature vector such as formula (1):
Wherein,It is storage state of the circulation layer in a upper time step, in t=1, value is unit matrix;
Finally, feature vector F is handled by S function, the visual angle finally predicted, the input viewing angle as future time step
vt+1, while the picture under the visual angle is obtained using RGBD camera, obtain the input tensor I of future time stept+1;
(1.3) training method:
The training of neural network is using enhancing learning method in the dynamic prediction module of visual angle, in each time step, according to formula
(2) reward is calculated:
Wherein,For the Three-dimension Reconstruction Model under t moment, V is the true model in database, and IoU is two model overlapping members
Plain number accounts for the specific gravity of whole elements;
After the reward for each time step that adds up, the optimization method based on strategy estimation, training network are used;Specifically used gradient
The gradient that the method for decline calculates then updates network parameter in the direction iteration that gradient reduces, obtains predicting most such as formula (3)
The neural network at excellent visual angle:
Wherein, R is the cumulative award of all time steps,The probability at the visual angle is obtained for prediction;
(2) Target self-determination rebuilds module:
(2.1) module inputs:
Picture is shot under the visual angle predicted in the dynamic prediction module of visual angle using RGBD camera, according to time step sequence by every frame
Picture obtains orderly picture tensor sequence { I for identical processing method in visual angle dynamic prediction module0, I1..., In,
Input as module;
(2.2) module architectures:
It is a Recognition with Recurrent Neural Network that Target self-determination, which rebuilds module, includes two parts of coding network and decoding network;Encode net
Network fencUsing the coding network of visual angle dynamic prediction module, i.e. a multilayer convolution loop neural network, it is responsible for the coding t time
The lower picture tensor I inputted of steptLow-dimensional feature Ft;Decoding network fdecIt is made of the three-dimensional deconvolution circulation layer of multilayer, it will be low
The picture feature F of dimensiontIt rises dimension and obtains three-dimensional voxelWhereinIt was circulation layer in the upper time
The storage state of step, in t=1, value is unit matrix;
The sequence of pictures being made of for one t time step is sequentially input in the dynamic prediction module of visual angle according to time step,
The three-dimensional voxel predicted under the last one time step is the three-dimensional reconstruction result of target;
(2.3) training method:
Target self-determination rebuild module using direction propagate with the training of the method for stochastic gradient descent, for a lot sample sheet, according to
Formula (4) calculates the error between the prediction result of network and database true value result, and calculates the gradient of error, according to nerve
The backpropagation of network gradually updates network parameter along the direction of gradient decline, and iteration is until convergence;
Lvox=| | Vpre-V||2 (4)
Wherein, VpreTarget self-determination, which is respectively indicated, with V rebuilds corresponding voxel in the voxel model and database of module network prediction
Model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810576919.3A CN108876907A (en) | 2018-05-31 | 2018-05-31 | A kind of active three-dimensional rebuilding method of object-oriented object |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810576919.3A CN108876907A (en) | 2018-05-31 | 2018-05-31 | A kind of active three-dimensional rebuilding method of object-oriented object |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108876907A true CN108876907A (en) | 2018-11-23 |
Family
ID=64336997
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810576919.3A Pending CN108876907A (en) | 2018-05-31 | 2018-05-31 | A kind of active three-dimensional rebuilding method of object-oriented object |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108876907A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109745062A (en) * | 2019-01-30 | 2019-05-14 | 腾讯科技(深圳)有限公司 | Generation method, device, equipment and the storage medium of CT image |
CN110458939A (en) * | 2019-07-24 | 2019-11-15 | 大连理工大学 | The indoor scene modeling method generated based on visual angle |
CN111127538A (en) * | 2019-12-17 | 2020-05-08 | 武汉大学 | Multi-view image three-dimensional reconstruction method based on convolution cyclic coding-decoding structure |
CN112950786A (en) * | 2021-03-01 | 2021-06-11 | 哈尔滨理工大学 | Vehicle three-dimensional reconstruction method based on neural network |
CN115115780A (en) * | 2022-06-29 | 2022-09-27 | 聚好看科技股份有限公司 | Three-dimensional reconstruction method and system based on multi-view RGBD camera |
CN115357845A (en) * | 2022-08-22 | 2022-11-18 | 浙江荷湖科技有限公司 | Evaluation method and device for microscopic light field iterative reconstruction result |
CN116630550A (en) * | 2023-07-21 | 2023-08-22 | 方心科技股份有限公司 | Three-dimensional model generation method and system based on multiple pictures |
US12016717B2 (en) | 2019-01-30 | 2024-06-25 | Tencent Technology (Shenzhen) Company Limited | CT image generation method and apparatus, computer device, and computer-readable storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9521399B1 (en) * | 2014-07-08 | 2016-12-13 | Aquifi, Inc. | Dynamically reconfigurable optical pattern generator module useable with a system to rapidly reconstruct three-dimensional data |
CN107330973A (en) * | 2017-07-03 | 2017-11-07 | 深圳市唯特视科技有限公司 | A kind of single-view method for reconstructing based on various visual angles supervision |
CN107862741A (en) * | 2017-12-10 | 2018-03-30 | 中国海洋大学 | A kind of single-frame images three-dimensional reconstruction apparatus and method based on deep learning |
CN108074218A (en) * | 2017-12-29 | 2018-05-25 | 清华大学 | Image super-resolution method and device based on optical field acquisition device |
-
2018
- 2018-05-31 CN CN201810576919.3A patent/CN108876907A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9521399B1 (en) * | 2014-07-08 | 2016-12-13 | Aquifi, Inc. | Dynamically reconfigurable optical pattern generator module useable with a system to rapidly reconstruct three-dimensional data |
CN107330973A (en) * | 2017-07-03 | 2017-11-07 | 深圳市唯特视科技有限公司 | A kind of single-view method for reconstructing based on various visual angles supervision |
CN107862741A (en) * | 2017-12-10 | 2018-03-30 | 中国海洋大学 | A kind of single-frame images three-dimensional reconstruction apparatus and method based on deep learning |
CN108074218A (en) * | 2017-12-29 | 2018-05-25 | 清华大学 | Image super-resolution method and device based on optical field acquisition device |
Non-Patent Citations (1)
Title |
---|
XIN YANG,ET AL: "Active Object Reconstruction Using a Guided View Planner", 《ARXIV:1805.03081V1》 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109745062A (en) * | 2019-01-30 | 2019-05-14 | 腾讯科技(深圳)有限公司 | Generation method, device, equipment and the storage medium of CT image |
WO2020156195A1 (en) * | 2019-01-30 | 2020-08-06 | 腾讯科技(深圳)有限公司 | Ct image generation method and apparatus, computer device and computer-readable storage medium |
US12016717B2 (en) | 2019-01-30 | 2024-06-25 | Tencent Technology (Shenzhen) Company Limited | CT image generation method and apparatus, computer device, and computer-readable storage medium |
CN110458939A (en) * | 2019-07-24 | 2019-11-15 | 大连理工大学 | The indoor scene modeling method generated based on visual angle |
CN110458939B (en) * | 2019-07-24 | 2022-11-18 | 大连理工大学 | Indoor scene modeling method based on visual angle generation |
CN111127538A (en) * | 2019-12-17 | 2020-05-08 | 武汉大学 | Multi-view image three-dimensional reconstruction method based on convolution cyclic coding-decoding structure |
CN111127538B (en) * | 2019-12-17 | 2022-06-07 | 武汉大学 | Multi-view image three-dimensional reconstruction method based on convolution cyclic coding-decoding structure |
CN112950786A (en) * | 2021-03-01 | 2021-06-11 | 哈尔滨理工大学 | Vehicle three-dimensional reconstruction method based on neural network |
CN115115780A (en) * | 2022-06-29 | 2022-09-27 | 聚好看科技股份有限公司 | Three-dimensional reconstruction method and system based on multi-view RGBD camera |
CN115357845A (en) * | 2022-08-22 | 2022-11-18 | 浙江荷湖科技有限公司 | Evaluation method and device for microscopic light field iterative reconstruction result |
CN116630550A (en) * | 2023-07-21 | 2023-08-22 | 方心科技股份有限公司 | Three-dimensional model generation method and system based on multiple pictures |
CN116630550B (en) * | 2023-07-21 | 2023-10-20 | 方心科技股份有限公司 | Three-dimensional model generation method and system based on multiple pictures |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108876907A (en) | A kind of active three-dimensional rebuilding method of object-oriented object | |
CN110458939B (en) | Indoor scene modeling method based on visual angle generation | |
Mou et al. | IM2HEIGHT: Height estimation from single monocular imagery via fully residual convolutional-deconvolutional network | |
Pumarola et al. | D-nerf: Neural radiance fields for dynamic scenes | |
CN105787439B (en) | A kind of depth image human synovial localization method based on convolutional neural networks | |
Kumar et al. | Monocular fisheye camera depth estimation using sparse lidar supervision | |
Zhang et al. | Deep learning-based classification and reconstruction of residential scenes from large-scale point clouds | |
CN105678284B (en) | A kind of fixed bit human body behavior analysis method | |
Hepp et al. | Learn-to-score: Efficient 3d scene exploration by predicting view utility | |
CN105205453B (en) | Human eye detection and localization method based on depth self-encoding encoder | |
Carvalho et al. | Multitask learning of height and semantics from aerial images | |
CN107945265A (en) | Real-time dense monocular SLAM method and systems based on on-line study depth prediction network | |
Wu et al. | Magicpony: Learning articulated 3d animals in the wild | |
CN112396703A (en) | Single-image three-dimensional point cloud model reconstruction method | |
CN103426200B (en) | Tree three-dimensional reconstruction method based on unmanned aerial vehicle aerial photo sequence image | |
Ulusoy et al. | Image-based 4-d reconstruction using 3-d change detection | |
CN108389226A (en) | A kind of unsupervised depth prediction approach based on convolutional neural networks and binocular parallax | |
CN110310285A (en) | A kind of burn surface area calculation method accurately rebuild based on 3 D human body | |
CN111160294A (en) | Gait recognition method based on graph convolution network | |
Wang et al. | Study on the method of transmission line foreign body detection based on deep learning | |
CN112734727A (en) | Apple picking method based on improved deep neural network | |
Condorelli et al. | A comparison between 3D reconstruction using nerf neural networks and mvs algorithms on cultural heritage images | |
CN115115685A (en) | Monocular image depth estimation algorithm based on self-attention neural network | |
Pirker et al. | Fast and accurate environment modeling using three-dimensional occupancy grids | |
Hirner et al. | FC-DCNN: A densely connected neural network for stereo estimation |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181123 |
|
WD01 | Invention patent application deemed withdrawn after publication |