CN110246212A - A kind of target three-dimensional rebuilding method based on self-supervisory study - Google Patents
A kind of target three-dimensional rebuilding method based on self-supervisory study Download PDFInfo
- Publication number
- CN110246212A CN110246212A CN201910367420.6A CN201910367420A CN110246212A CN 110246212 A CN110246212 A CN 110246212A CN 201910367420 A CN201910367420 A CN 201910367420A CN 110246212 A CN110246212 A CN 110246212A
- Authority
- CN
- China
- Prior art keywords
- binary map
- point cloud
- image
- self
- 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.)
- Granted
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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/90—Determination of colour characteristics
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10028—Range image; Depth image; 3D point clouds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Graphics (AREA)
- Geometry (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to a kind of target three-dimensional rebuilding methods based on self-supervisory study, comprising: S1, training points cloud autoencoder network;S2, training binary map autoencoder network;S3, input RGB image, obtain true binary map;S4, image pose is extracted using Pose net;S5, training image encoder, and generate preliminary point cloud model;S6, transformation point cloud model is generated;S7, training points cloud encoder, and generate recovery binary map;S8, the mean square deviation for restoring binary map and true binary map is calculated, if mean square deviation is less than preset threshold, exported as a result, no then follow the steps S9;S9, feedback mean square deviation are to image encoder, and return step S5 again.Compared with prior art, the present invention using Pose net extraction image pose and increases two-dimentional supervision message, solves the problems such as fuzzy input picture visual angle, shortage supervision item information, improves the accuracy of target three-dimensional reconstruction.
Description
Technical field
The present invention relates to computer vision and technical field of image processing, more particularly, to a kind of based on self-supervisory study
Target three-dimensional rebuilding method.
Background technique
As a research direction of computer vision and computer graphics height intersection, three-dimensional reconstruction passes through specific
Device and algorithm rebuild the mathematical model of three-dimension object in the real world, have been widely used for Intelligent unattended
Multiple industries such as system, robot, computer aided medicine, virtual reality, augmented reality.Traditional three-dimensional rebuilding method is ground
Study carefully and focus mostly on multi-view geometry, including SFM and SLAM, although these methods all achieve certain effect in special scenes
Fruit, but there is also some drawbacks: 1) part that multi-view geometry can not lack in reconstructed view needs to input enough views
To guarantee the integrality of reconstructed object;2) increase for meaning computation complexity to the reconstruction of multiple view is difficult to accomplish weight in real time
The effect built.These drawbacks all limit the application of multiple view reconstruction, and therefore, the method based on study realizes the reconstruction of single-view
It is particularly important.
Mainly include two methods currently based on the single-view image reconstruction of study: a kind of method is generated according to 2D CNN
The method of image is generated by 3D CNN, and with voxel shape representation;Another method is to pass through change based on Weakly supervised mechanism
Constituent encoder fits 3D shape, and 3D-RecGAN is the new network of one kind for being suggested based on GAN structure, can directly from
The voxel architecture of individual depth map reply object.The voxel that above two method obtains is 256^3, the requirement to hardware device
It is very high, since the 3D shape that voxel indicates is to improve surface accuracy by increasing the fast number of 3 D stereo, in order to
EQUILIBRIUM CALCULATION FOR PROCESS complexity and surface accuracy have some method for reconstructing based on grid and point cloud to be suggested again in the recent period, for example,
Individual color image directly can be generated triangle gridding by the thought of Pixel2Mesh combination picture scroll product;PSG-Net utilizes unordered
The network frame and loss function of point cloud realize the point Yun Chongjian indicated better than voxel.But since single-view reconstruction itself lacks
Lack enough supervision item information and part input picture visual angle is fuzzy, causes the model reconstructed there are excalation, lacks carefully
Situations such as saving surface abundant.
Summary of the invention
It is an object of the present invention to overcome the above-mentioned drawbacks of the prior art and provide one kind to be based on self-supervisory
The target three-dimensional rebuilding method of habit, to improve the accuracy of target three-dimensional reconstruction.
The purpose of the present invention can be achieved through the following technical solutions: a kind of target Three-dimensional Gravity based on self-supervisory study
Construction method, comprising the following steps:
S1, training points cloud autoencoder network obtain the point potential feature of cloudWherein, point cloud autoencoder network includes point cloud
Self-encoding encoderWith a cloud decoder DP;
S2, training binary map autoencoder network, obtain the potential feature of binary mapWherein, binary map autoencoder network packet
Include binary map self-encoding encoderWith binary map decoder DI;
S3, input RGB image, obtain true binary map by binary conversion treatment;
S4, feature extraction, the image position of the RGB image inputted are carried out using RGB image of the Pose net to input
Appearance;
S5, the RGB image according to input, training image encoderObtain the first space characteristics FP, pass through the step
Point cloud decoder D in S1P, the RGB image of input is generated into preliminary point cloud model;
S6, translation and rotation transformation are carried out to preliminary point cloud model by image pose, generates transformation point cloud model;
S7, according to transformation point cloud model, training points cloud encoderObtain second space feature FB, pass through the step
Binary map decoder D in S2I, generate and restore binary map;
S8, the mean square deviation for restoring binary map and true binary map is calculated, if the mean square deviation is less than preset threshold,
Output transform point cloud model is the target three-dimensional reconstruction result for inputting RGB image, no to then follow the steps S9;
S9, feedback mean square deviation are to image encoderAnd return step S5 again.
Preferably, the point potential feature of cloud is obtained in the step S1Detailed process are as follows:
S11, by true point cloud data input point cloud self-encoding encoderB × N × 512 are obtained after 5 layer of 1 dimension convolution
Feature;
The feature of B × N × 512 obtains the point Yun Qian of B × 512 by maximum pond layer operation in S12, the step S11
In featureWherein, k=512.
Preferably, the midpoint step S1 cloud decoder DPIncluding three layers of full articulamentum, the potential feature of cloud will be putTurn
Change the point cloud format of B × N × 3 into.
Preferably, the potential feature of binary map is obtained in the step S2Detailed process are as follows:
S21, binary conversion treatment is carried out to RGB image, obtains true binary mapWherein, binary conversion treatment is by pixel
The place that value is 0 indicates that pixel value is that the place of non-zero is indicated with 1 with 0;
S22, by true binary mapInput binary map self-encoding encoderObtain the potential feature of binary map Its
In, k=512.
Preferably, binary map decoder D in the step S2IIt is operated using deconvolution and carries out picture material filling, so that
Picture material is gradually enriched to restore binary map.
Preferably, Pose net uses full articulamentum to return out image aspects information in the step S4, to obtain image
Pose, described image Viewing-angle information include (α, beta, gamma, a, b, c) totally six parameters, wherein (α, beta, gamma) is deflection, respectively
Indicate that yaw angle, pitch angle and roll angle, (a, b, c) are translation vector;
Described image pose is (R, t), wherein R is spin matrix, t=(a, b, c), by deflection (α, beta, gamma) to rotation
The conversion formula of torque battle array R is as follows:
Preferably, the step S5 specifically includes the following steps:
S51, input RGB image are to image encoderObtain the first space characteristics FP;
S52, by the first space characteristics FPWith the potential feature of cloudConstitute first-loss function;
S53, according to first-loss function and the first space characteristics FP, using a cloud decoder DPGenerate preliminary point cloud model.
Preferably, it is specifically by image pose and preliminary point cloud model phase that transformation point cloud model is generated in the step S6
Multiply, so that preliminary point cloud model be made to transform to camera plane:
x′i=Rxi+t i∈[0,N-1]
Wherein, xiFor the point in preliminary point cloud model, x 'iFor the point in transformation point cloud model, N indicates to wrap in three-dimensional structure
The number of the point contained, each point xiIt is multiplied after transformation with image pose (R, t) and obtains x 'i。
Preferably, the step S7 specifically includes the following steps:
S71, point cloud model input point cloud encoder will be convertedObtain second space feature FB;
S72, by second space feature FBWith the potential feature of binary mapConstitute the second loss function;
S73, according to the second loss function and second space feature FB, using binary map decoder DIIt generates and restores binary map.
Compared with prior art, the invention has the following advantages that
One, the present invention, which carries out feature extraction to RGB image using Pose net, has network to obtain image pose
The ability at resolution image visual angle solves the problems, such as the input picture dimness of vision, can generate reasonable point cloud by operative constraint
Model.
Two, the present invention is based on image poses translates preliminary point cloud model, rotation transformation, and is generated by network extensive
Multiple binary map, to make full use of binary map information to carry out self-supervisory, improve mesh as the supervision item information for generating point cloud model
Mark the accuracy of three-dimensional reconstruction result.
Detailed description of the invention
Fig. 1 is method flow schematic diagram of the invention;
Fig. 2 is the target three-dimensional reconstruction process schematic of embodiment;
Fig. 3 is the schematic network structure of Pose net.
Specific embodiment
The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.
As shown in Figure 1, a kind of target three-dimensional rebuilding method based on self-supervisory study, comprising the following steps:
S1, training points cloud autoencoder network obtain the point potential feature of cloudWherein, point cloud autoencoder network includes point cloud
Self-encoding encoderWith a cloud decoder DP;
S2, training binary map autoencoder network, obtain the potential feature of binary mapWherein, binary map autoencoder network packet
Include binary map self-encoding encoderWith binary map decoder DI;
S3, input RGB image, obtain true binary map by binary conversion treatment;
S4, feature extraction, the image position of the RGB image inputted are carried out using RGB image of the Pose net to input
Appearance;
S5, the RGB image according to input, training image encoderObtain the first space characteristics FP, pass through the step
Point cloud decoder D in S1P, the RGB image of input is generated into preliminary point cloud model;
S6, translation and rotation transformation are carried out to preliminary point cloud model by image pose, generates transformation point cloud model;
S7, according to transformation point cloud model, training points cloud encoderObtain second space feature FB, pass through the step
Binary map decoder D in S2I, generate and restore binary map;
S8, the mean square deviation for restoring binary map and true binary map is calculated, if the mean square deviation is less than preset threshold,
Output transform point cloud model is the target three-dimensional reconstruction result for inputting RGB image, no to then follow the steps S9;
S9, feedback mean square deviation are to image encoderAnd return step S5 again.
In the present embodiment, cloud self-encoding encoder is putBinary map self-encoding encoderImage encoderPoint cloud encoderPoint cloud decoder DPWith binary map decoder DINetwork structure it is as shown in table 1:
Table 1
Wherein, the detailed process of step S1 includes: to be sent into point cloud certainly with the true point cloud data B × X × 3 of aircraft for one group
EncoderThe feature of B × N × 512 is obtained after 5 layer of 1 dimension convolution, then obtains B × 512 after being operated by Max pooling
The potential feature of point cloudPoint cloud decoder DPPart is made of three layers of full articulamentum, will finally put cloud
Potential featureIt is deformed into the point cloud format of B × N × 3, obtains aircraft point cloud model;
The detailed process of step S2 includes: to carry out binary conversion treatment to the RGB image of one group of aircraft, i.e. pixel value is 0
Place indicates that pixel value is that the place of non-zero is indicated with 1 with 0, and to obtain true binary map, true binary map is then sent into two
It is worth figure self-encoding encoderObtain the potential feature of binary mapBinary map decoder DIIt is grasped using deconvolution
Make filling picture material, so that picture material becomes gradually to enrich to recover the binary map of aircraft RGB image.
The present embodiment is to carry out target three-dimensional reconstruction to aircraft brake disc, and process schematic is as shown in Figure 2.Wherein, step S3
It is that binary conversion treatment is carried out to the aircraft RGB image of input, obtains the true binary map of aircraft;
The detailed process of step S4 includes: to obtain the image pose (R, t) of input RGB image, Pose by Pose net
The network structure of net indicates image aspects letter as shown in figure 3, its last layer returns out a six-vector using full articulamentum
Breath, image aspects information include that (a, beta, gamma, a, b, c) totally six parameters, (α, beta, gamma) respectively indicate three deflections, that is, yaw
Angle, pitch angle and roll angle, (a, b, c) indicates translation vector, and in image pose (R, t), R is spin matrix, t=(a, b, c),
And the conversion formula of deflection to spin matrix R are as follows:
Image encoder needed for generating preliminary point cloud model in step S5With binary map self-encoding encoderNetwork knot
Structure is identical, i.e., in addition to the last layer is full articulamentum, remaining network layer is convolutional layer, thus extracts to the RGB image of input
Further feature simultaneously obtains the first space characteristics of aircraft F of 512 dimensions finally by full articulamentumP, and by the first space characteristics FPAnd point
The potential feature of cloudIt calculates mean square error and constitutes L1 normalization loss function, decoder directlys adopt trained point in step S1
Cloud decoder DP, to obtain preliminary aircraft point cloud model;
The detailed process of step S6 includes: preliminary to generating in step S5 using the image pose (R, t) in step S4
Point cloud model carries out translation and rotation transformation, i.e., is multiplied with first beans-and bullets shooter cloud, so that preliminary point cloud model is transformed to camera plane, obtain
The transformation point cloud model of aircraft makes the 3D aircraft shape of the corresponding pose of a binary map, to realize a cloud and binary map
One-to-one correspondence:
x′i=Rxi+t i∈[0,N-1]
In formula, N indicates the number for the point for including in three-dimensional structure, each of preliminary point cloud model point xiBy matrix
X ' is obtained after (R, t) transformationi, x 'iFor the point in transformation point cloud model;
Step S7 is that the transformation point cloud model of aircraft is sent into point cloud encoderIt performs the encoding operation, exports aircraft second
Space characteristics FB, and with the potential feature of binary mapIt constitutes L1 and normalizes loss function, decoder directlys adopt instructs in step S2
The binary map decoder D perfectedI, to obtain the recovery binary map of aircraft;
The detailed process of step S8 and S9 include: the mean square error differential loss of the recovery binary map and true binary map that calculate aircraft
Mistake value, if the mean square error penalty values are less than preset threshold, it is winged for exporting the aircraft transformation point cloud model generated in step S6
Otherwise the mean square error penalty values are fed back to image encoder by the target three-dimensional reconstruction result of machineRecirculate step S5
To step S9, until optimal aircraft point cloud model is obtained, using the target three-dimensional reconstruction result as aircraft.
Claims (9)
1. a kind of target three-dimensional rebuilding method based on self-supervisory study, which comprises the following steps:
S1, training points cloud autoencoder network obtain the point potential feature of cloudWherein, point cloud autoencoder network includes that point cloud is self-editing
Code deviceWith a cloud decoder DP;
S2, training binary map autoencoder network, obtain the potential feature of binary mapWherein, binary map autoencoder network includes two
It is worth figure self-encoding encoderWith binary map decoder DI;
S3, input RGB image, obtain true binary map by binary conversion treatment;
S4, feature extraction, the image pose of the RGB image inputted are carried out using RGB image of the Pose net to input;
S5, the RGB image according to input, training image encoderObtain the first space characteristics FP, by the step S1
Point cloud decoder DP, the RGB image of input is generated into preliminary point cloud model;
S6, translation and rotation transformation are carried out to preliminary point cloud model by image pose, generates transformation point cloud model;
S7, according to transformation point cloud model, training points cloud encoderObtain second space feature FB, by the step S2
Binary map decoder DI, generate and restore binary map;
S8, the mean square deviation for restoring binary map and true binary map is calculated, if the mean square deviation is less than preset threshold, exported
Transformation point cloud model is the target three-dimensional reconstruction result for inputting RGB image, no to then follow the steps S9;
S9, feedback mean square deviation are to image encoderAnd return step S5 again.
2. a kind of target three-dimensional rebuilding method based on self-supervisory study according to claim 1, which is characterized in that described
The point potential feature of cloud is obtained in step S1Detailed process are as follows:
S11, by true point cloud data input point cloud self-encoding encoderThe feature of B × N × 512 is obtained after 5 layer of 1 dimension convolution;
The feature of B × N × 512 obtains the potential spy of point cloud of B × 512 by maximum pond layer operation in S12, the step S11
SignWherein, k=512.
3. a kind of target three-dimensional rebuilding method based on self-supervisory study according to claim 1, which is characterized in that described
The midpoint step S1 cloud decoder DPIncluding three layers of full articulamentum, the potential feature of cloud will be putIt is converted into the point cloud lattice of B × N × 3
Formula.
4. a kind of target three-dimensional rebuilding method based on self-supervisory study according to claim 1, which is characterized in that described
The potential feature of binary map is obtained in step S2Detailed process are as follows:
S21, binary conversion treatment is carried out to RGB image, obtains true binary mapWherein, it is 0 that binary conversion treatment, which is by pixel value,
Place indicated with 0, pixel value be non-zero place indicated with 1;
S22, by true binary mapInput binary map self-encoding encoderObtain the potential feature of binary map Wherein, k
=512.
5. a kind of target three-dimensional rebuilding method based on self-supervisory study according to claim 1, which is characterized in that described
Binary map decoder D in step S2IUsing deconvolution operate carry out picture material filling so that picture material gradually enrich to
Restore binary map.
6. a kind of target three-dimensional rebuilding method based on self-supervisory study according to claim 1, which is characterized in that described
Pose net returns out image aspects information using full articulamentum in step S4, to obtain image pose, described image visual angle letter
Breath includes (α, beta, gamma, a, b, c) totally six parameters, wherein (α, beta, gamma) is deflection, respectively indicates yaw angle, pitch angle and
Roll angle, (a, b, c) are translation vector;
Described image pose is (R, t), wherein R is spin matrix, and t=(a, b, c) arrives spin moment by deflection (α, beta, gamma)
The conversion formula of battle array R is as follows:
7. a kind of target three-dimensional rebuilding method based on self-supervisory study according to claim 1, which is characterized in that described
Step S5 specifically includes the following steps:
S51, input RGB image are to image encoderObtain the first space characteristics FP;
S52, by the first space characteristics FPWith the potential feature of cloudConstitute first-loss function;
S53, according to first-loss function and the first space characteristics FP, using a cloud decoder DPGenerate preliminary point cloud model.
8. a kind of target three-dimensional rebuilding method based on self-supervisory study according to claim 1, which is characterized in that described
Step S7 specifically includes the following steps:
S71, point cloud model input point cloud encoder will be convertedObtain second space feature FB;
S72, by second space feature FBWith the potential feature of binary mapConstitute the second loss function;
S73, according to the second loss function and second space feature FB, using binary map decoder DIIt generates and restores binary map.
9. a kind of target three-dimensional rebuilding method based on self-supervisory study according to claim 6, which is characterized in that described
It is specifically that image pose is multiplied with preliminary point cloud model that transformation point cloud model is generated in step S6, to make preliminary point cloud model
Transform to camera plane:
x′i=Rxi+t i∈[0,N-1]
Wherein, xiFor the point in preliminary point cloud model, x 'iTo convert the point in point cloud model, include in N expression three-dimensional structure
The number of point, each point xiIt is multiplied after transformation with image pose (R, t) and obtains x 'i。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910367420.6A CN110246212B (en) | 2019-05-05 | 2019-05-05 | Target three-dimensional reconstruction method based on self-supervision learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910367420.6A CN110246212B (en) | 2019-05-05 | 2019-05-05 | Target three-dimensional reconstruction method based on self-supervision learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110246212A true CN110246212A (en) | 2019-09-17 |
CN110246212B CN110246212B (en) | 2023-02-07 |
Family
ID=67883786
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910367420.6A Active CN110246212B (en) | 2019-05-05 | 2019-05-05 | Target three-dimensional reconstruction method based on self-supervision learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110246212B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111311664A (en) * | 2020-03-03 | 2020-06-19 | 上海交通大学 | Joint unsupervised estimation method and system for depth, pose and scene stream |
CN112767468A (en) * | 2021-02-05 | 2021-05-07 | 中国科学院深圳先进技术研究院 | Self-supervision three-dimensional reconstruction method and system based on collaborative segmentation and data enhancement |
CN113438481A (en) * | 2020-03-23 | 2021-09-24 | 富士通株式会社 | Training method, image coding method, image decoding method and device |
CN113592913A (en) * | 2021-08-09 | 2021-11-02 | 中国科学院深圳先进技术研究院 | Method for eliminating uncertainty of self-supervision three-dimensional reconstruction |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1747559A (en) * | 2005-07-29 | 2006-03-15 | 北京大学 | Three-dimensional geometric mode building system and method |
US20100309304A1 (en) * | 2007-10-24 | 2010-12-09 | Bernard Chalmond | Method and Device for the Reconstruction of the Shape of an Object from a Sequence of Sectional Images of Said Object |
CN107481313A (en) * | 2017-08-18 | 2017-12-15 | 深圳市唯特视科技有限公司 | A kind of dense three-dimensional object reconstruction method based on study available point cloud generation |
CN108665499A (en) * | 2018-05-04 | 2018-10-16 | 北京航空航天大学 | A kind of low coverage aircraft pose measuring method based on parallax method |
CN108694741A (en) * | 2017-04-07 | 2018-10-23 | 杭州海康威视数字技术股份有限公司 | A kind of three-dimensional rebuilding method and device |
CN109087325A (en) * | 2018-07-20 | 2018-12-25 | 成都指码科技有限公司 | A kind of direct method point cloud three-dimensional reconstruction and scale based on monocular vision determines method |
CN109583304A (en) * | 2018-10-23 | 2019-04-05 | 宁波盈芯信息科技有限公司 | A kind of quick 3D face point cloud generation method and device based on structure optical mode group |
-
2019
- 2019-05-05 CN CN201910367420.6A patent/CN110246212B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1747559A (en) * | 2005-07-29 | 2006-03-15 | 北京大学 | Three-dimensional geometric mode building system and method |
US20100309304A1 (en) * | 2007-10-24 | 2010-12-09 | Bernard Chalmond | Method and Device for the Reconstruction of the Shape of an Object from a Sequence of Sectional Images of Said Object |
CN108694741A (en) * | 2017-04-07 | 2018-10-23 | 杭州海康威视数字技术股份有限公司 | A kind of three-dimensional rebuilding method and device |
CN107481313A (en) * | 2017-08-18 | 2017-12-15 | 深圳市唯特视科技有限公司 | A kind of dense three-dimensional object reconstruction method based on study available point cloud generation |
CN108665499A (en) * | 2018-05-04 | 2018-10-16 | 北京航空航天大学 | A kind of low coverage aircraft pose measuring method based on parallax method |
CN109087325A (en) * | 2018-07-20 | 2018-12-25 | 成都指码科技有限公司 | A kind of direct method point cloud three-dimensional reconstruction and scale based on monocular vision determines method |
CN109583304A (en) * | 2018-10-23 | 2019-04-05 | 宁波盈芯信息科技有限公司 | A kind of quick 3D face point cloud generation method and device based on structure optical mode group |
Non-Patent Citations (3)
Title |
---|
ANDRÉ HENN, GERHARDGRÖGER, VIKTOR STROH,LUTZ PLÜMER: ""Model driven reconstruction of roofs from sparse LIDAR point clouds"", 《ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING》 * |
THOMAS KRIJNEN,JAKO BBEETZ: ""An IFC schema extension and binary serialization format to efficiently integrate point cloud data into building models"", 《ADVANCED ENGINEERING INFORMATICS》 * |
黄煜: ""基于靶标特征的机器人视觉自主寻位方法"", 《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 |
---|---|---|---|---|
CN111311664A (en) * | 2020-03-03 | 2020-06-19 | 上海交通大学 | Joint unsupervised estimation method and system for depth, pose and scene stream |
CN111311664B (en) * | 2020-03-03 | 2023-04-21 | 上海交通大学 | Combined unsupervised estimation method and system for depth, pose and scene flow |
CN113438481A (en) * | 2020-03-23 | 2021-09-24 | 富士通株式会社 | Training method, image coding method, image decoding method and device |
CN113438481B (en) * | 2020-03-23 | 2024-04-12 | 富士通株式会社 | Training method, image encoding method, image decoding method and device |
CN112767468A (en) * | 2021-02-05 | 2021-05-07 | 中国科学院深圳先进技术研究院 | Self-supervision three-dimensional reconstruction method and system based on collaborative segmentation and data enhancement |
WO2022166412A1 (en) * | 2021-02-05 | 2022-08-11 | 中国科学院深圳先进技术研究院 | Self-supervised three-dimensional reconstruction method and system based on collaborative segmentation and data enhancement |
CN112767468B (en) * | 2021-02-05 | 2023-11-03 | 中国科学院深圳先进技术研究院 | Self-supervision three-dimensional reconstruction method and system based on collaborative segmentation and data enhancement |
CN113592913A (en) * | 2021-08-09 | 2021-11-02 | 中国科学院深圳先进技术研究院 | Method for eliminating uncertainty of self-supervision three-dimensional reconstruction |
CN113592913B (en) * | 2021-08-09 | 2023-12-26 | 中国科学院深圳先进技术研究院 | Method for eliminating uncertainty of self-supervision three-dimensional reconstruction |
Also Published As
Publication number | Publication date |
---|---|
CN110246212B (en) | 2023-02-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110246212A (en) | A kind of target three-dimensional rebuilding method based on self-supervisory study | |
Mandikal et al. | Dense 3d point cloud reconstruction using a deep pyramid network | |
Le et al. | Pointgrid: A deep network for 3d shape understanding | |
CN110288695B (en) | Single-frame image three-dimensional model surface reconstruction method based on deep learning | |
CN102592275B (en) | Virtual viewpoint rendering method | |
CN111862101A (en) | 3D point cloud semantic segmentation method under aerial view coding visual angle | |
CN112396703A (en) | Single-image three-dimensional point cloud model reconstruction method | |
Zhang et al. | Research on image processing technology of computer vision algorithm | |
CN108280858A (en) | A kind of linear global camera motion method for parameter estimation in multiple view reconstruction | |
CN110827295A (en) | Three-dimensional semantic segmentation method based on coupling of voxel model and color information | |
CN107993255A (en) | A kind of dense optical flow method of estimation based on convolutional neural networks | |
CN102306386A (en) | Method for quickly constructing third dimension tree model from single tree image | |
CN110889901B (en) | Large-scene sparse point cloud BA optimization method based on distributed system | |
CN114463511A (en) | 3D human body model reconstruction method based on Transformer decoder | |
CN110443883A (en) | A kind of individual color image plane three-dimensional method for reconstructing based on dropblock | |
CN113077554A (en) | Three-dimensional structured model reconstruction method based on any visual angle picture | |
CN106127743B (en) | The method and system of automatic Reconstruction bidimensional image and threedimensional model accurate relative location | |
CN115951784A (en) | Dressing human body motion capture and generation method based on double nerve radiation fields | |
CN106251281A (en) | A kind of image morphing method based on shape interpolation | |
CN114677479A (en) | Natural landscape multi-view three-dimensional reconstruction method based on deep learning | |
CN116134491A (en) | Multi-view neuro-human prediction using implicit differentiable renderers for facial expression, body posture morphology, and clothing performance capture | |
Liu et al. | Facial image inpainting using attention-based multi-level generative network | |
Fang et al. | One is all: Bridging the gap between neural radiance fields architectures with progressive volume distillation | |
CN108230431B (en) | Human body action animation generation method and system of two-dimensional virtual image | |
CN103971397B (en) | The global illumination method for drafting reduced based on virtual point source and sparse matrix |
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 |