CN114359510A - Point cloud completion method based on anchor point detection - Google Patents
Point cloud completion method based on anchor point detection Download PDFInfo
- Publication number
- CN114359510A CN114359510A CN202111504436.0A CN202111504436A CN114359510A CN 114359510 A CN114359510 A CN 114359510A CN 202111504436 A CN202111504436 A CN 202111504436A CN 114359510 A CN114359510 A CN 114359510A
- Authority
- CN
- China
- Prior art keywords
- point cloud
- point
- anchor
- points
- method based
- 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
Images
Abstract
The invention belongs to the technical field of point cloud completion, and relates to a point cloud completion method based on anchor point detection, which comprises the steps of firstly preprocessing a data set by using a local deletion operation so as to simulate an incomplete point cloud; then dividing the space where the point cloud is located into a plurality of anchor points, training a neural network to judge the positivity and negativity of the anchor points, and obtaining positive sample anchor points with uniform density and regular shape; and finally, generating dense point cloud in the neighborhood of the positive sample anchor point through a neural network, realizing the diffusion from the single anchor point to the local part, and completing the point cloud completion. In order to realize the method, a GCNet model matched with the method is trained, the neural network model can judge the positivity and the negativity of the anchor points and generate three-dimensional coordinate information of anchor point diffusion, end-to-end point cloud completion is realized, the problem of poor completion effect caused by the fact that the existing method only depends on neural network regression to generate point cloud coordinates is solved, and the point cloud completion effect is improved.
Description
Technical Field
The invention belongs to the field of point cloud completion, and particularly relates to a point cloud completion method based on anchor point detection.
Background
Predicting the complete 3D point cloud is a core task of many computer vision.
The existing method is usually based on deep learning, and global features are extracted from incomplete point clouds and used for restoring the complete point clouds. Although they predict complete point clouds to some extent, the structural details of the object cannot be accurately generated, and the problem of uneven density exists.
Applying deep learning techniques to point cloud data is a direction of recent growth. For the point cloud completion task, a neural network can be trained to achieve the purpose, namely: inputting an incomplete point cloud structure and generating a complete point cloud structure.
At present, most network models based on deep learning have poor effect on point cloud completion because the models simply classify the task of point cloud completion into a regression problem and train a neural network to predict point cloud coordinates, so that only point clouds which are substantially correct and irregular in shape can be generated. Therefore, it is necessary to realize point cloud completion with regular shape and uniform density.
Disclosure of Invention
In order to overcome the defects of local feature extraction in the prior art, the invention provides a point cloud complementing method based on anchor point detection, and a network structure matched with the point cloud complementing method is designed, so that regular and uniform point cloud coordinates can be generated. Experiments prove that the invention has excellent effect on vision and indexes.
The technical scheme adopted by the invention is as follows:
a point cloud completion method based on anchor point detection comprises the following steps:
s1, data set acquisition and data preprocessing: acquiring a point cloud public data set ShapeNet, and deleting local point clouds in complete point cloud data by using local deletion operation so as to construct incomplete point clouds;
s2, constructing a neural network model: based on the point cloud completion thought from rough to fine, firstly, judging through the positive and negative properties of an anchor point to obtain sparse point cloud with regular shape and uniform density, then obtaining dense point cloud through the diffusion of the sparse point cloud, and constructing a GCNet network model for realizing the thought;
s3, model training and model saving: and (3) training the network by using an Adam optimizer, reducing a loss function, improving a point cloud completion effect, and storing the model when the loss function of the model tends to be stable.
In step S1, the public data set sharenet includes 16 kinds of point cloud data, each individual point cloud data file (.pts) of all kinds is composed of n point cloud coordinates x, y, z in a three-dimensional coordinate system, and the value of n is different.
In step S1, the specific method of the local deletion operation includes definition and selection of a key point and deletion of a point cloud adjacent to the key point.
In step S1, the specific method for defining and selecting the key points includes: taking 4 points of (1,0,0), (0,0,1), (1,0,1) and (-1,0,0) as key points, wherein three numbers in (b) represent coordinates of x, y and z in a three-dimensional coordinate system; in order to simulate incomplete point clouds with different parts missing, 4 key points are sequentially used as central points, and point cloud data near the central point are deleted; the method comprises the following specific steps of: and calculating the distances between all points in the point cloud and the center, and deleting 512 point clouds closest to the center point.
In the step S2, in the sparse point cloud with regular shape and uniform density obtained through the positive and negative judgment of the anchor points, since the number of negative samples in the space is far greater than that of positive samples and excessive negative samples have no benefit for network training, in order to solve the problem of excessive negative samples, only three times of the number of positive samples are selected from the negative samples for training. In addition, in the selected negative samples, the negative samples are further divided into difficultly trained negative samples and easily trained negative samples by setting a threshold value d, and the difficultly trained negative samples with the distance smaller than the threshold value d are selected; the number ratio of the negative samples which are difficult to train to the negative samples which are easy to train is 2: 1.
in the step S2, in the dense point cloud obtained by the diffusion of the sparse point cloud, since the sparse point cloud can only represent the shape structure of the point cloud as a whole, but cannot display the specific details thereof, point cloud coordinates in the neighborhood of each anchor point are generated by a neural network on the basis of the sparse point cloud (anchor point sample), and the maximum range of the neighborhood is limited by tanh, so that the maximum range of the neighborhood does not exceed half of the distance between two anchor points.
In step S2, the GCNet includes two parts, namely an encoder part and a decoder part, and the specific goals of the encoder part are: extracting the coordinate characteristics of the input incomplete point cloud; the specific structure of the encoder part is as follows: firstly, inputting point cloud coordinates, obtaining a feature vector of the point cloud through a layer of convolution layer, then dividing a space where the point cloud is located into w multiplied by w anchor points, averaging vectors in grids corresponding to the anchor points so as to realize grid structurization of features, and finally extracting high-dimensional features through 3D convolution and residual connection; the specific goals of the decoder part are: generating point cloud coordinates of the missing part through the high-dimensional features extracted by the decoder part; the specific structure of the decoder part is as follows: the input high-dimensional features are subjected to 3D transposition convolution to realize feature up-sampling, and the feature up-sampling is used for predicting the positive and negative properties of anchor points and the coordinate information of the diffused point cloud in the neighborhood of the positive anchor points to generate the final complete point cloud.
In step S3, the loss function includes BCELoss and CD, and the specific formula of BCELoss is:
wherein XiTo predict value, yiFor labels, the specific formula for a CD is:
wherein CD (S)1,S2) Point cloud set S representing predictive generation1And a real point cloud set S2The Euclidean space average closest distance between; and x and y respectively represent one point cloud in the point cloud set and consist of three coordinates of x, y and z.
Compared with the prior art, the invention has the following beneficial effects:
the invention abandons the idea of predicting the point cloud coordinate by a pure regression idea, trains the neural network GCNet, can judge the positivity of the anchor point and generate three-dimensional coordinate information of anchor point diffusion by the neural network model, realizes end-to-end point cloud completion, and solves the problem of poor completion effect caused by the existing method which only depends on neural network regression to generate the point cloud coordinate. The point cloud complementing effect is improved.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic diagram of the GCNet structure of the present invention;
FIG. 3 is a point cloud completion effect of the airplane in example 1;
FIG. 4 shows the point cloud completion effect of the airplane in example 2;
FIG. 5 shows the point cloud completion effect of the airplane in example 3;
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described below, and it is obvious that the described embodiments are a part of the embodiments of the present invention, but not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
A point cloud completion method based on anchor point detection uses a Shapelet data set to conduct point cloud completion training, wherein the data set comprises 16 types and has 15011 point cloud data, 12137 point cloud data are trained, and 2874 point cloud data are tested. And preprocessing each point cloud data through viewpoint operation to obtain incomplete point clouds input by a network. Inputting the incomplete point cloud into GCNet, training by using BCELoss and CD loss functions, and storing the weight with the best effect after 100 rounds of training. As shown in fig. 1, the detailed process includes the following steps:
s1, data set acquisition and data preprocessing: acquiring a point cloud public data set ShapeNet, and deleting local point clouds in complete point cloud data by using local deletion operation so as to construct incomplete point clouds;
s2, constructing a neural network model: based on the point cloud completion thought from rough to fine, firstly, judging through the positive and negative properties of an anchor point to obtain sparse point cloud with regular shape and uniform density, then obtaining dense point cloud through the diffusion of the sparse point cloud, and constructing a GCNet network model for realizing the thought;
s3, model training and model saving: and (3) training the network by using an Adam optimizer, reducing a loss function, improving a point cloud completion effect, and storing the model when the loss function of the model tends to be stable.
In step S1, the public data set sharenet includes 16 kinds of point cloud data, each individual point cloud data file (.pts) of all kinds is composed of n point cloud coordinates x, y, z in a three-dimensional coordinate system, and the value of n is different.
In step S1, the specific method of the local deletion operation includes definition and selection of a key point and deletion of a point cloud adjacent to the key point.
In step S1, the specific method for defining and selecting the key points includes: taking 4 points of (1,0,0), (0,0,1), (1,0,1) and (-1,0,0) as key points, wherein three numbers in (b) represent coordinates of x, y and z in a three-dimensional coordinate system; in order to simulate incomplete point clouds with different parts missing, 4 key points are sequentially used as central points, and point cloud data near the central point are deleted; the method comprises the following specific steps of: and calculating the distances between all points in the point cloud and the center, and deleting 512 point clouds closest to the center point.
In the step S2, in the sparse point cloud with regular shape and uniform density obtained through the positive and negative judgment of the anchor points, since the number of negative samples in the space is far greater than that of positive samples and excessive negative samples have no benefit for network training, in order to solve the problem of excessive negative samples, only three times of the number of positive samples are selected from the negative samples for training. In addition, in the selected negative samples, the negative samples are further divided into difficultly trained negative samples and easily trained negative samples by setting a threshold value d, and the difficultly trained negative samples with the distance smaller than the threshold value d are selected; the number ratio of the negative samples which are difficult to train to the negative samples which are easy to train is 2: 1.
in the step S2, in the dense point cloud obtained by the diffusion of the sparse point cloud, since the sparse point cloud can only represent the shape structure of the point cloud as a whole, but cannot display the specific details thereof, point cloud coordinates in the neighborhood of each anchor point are generated by a neural network on the basis of the sparse point cloud (anchor point sample), and the maximum range of the neighborhood is limited by tanh, so that the maximum range of the neighborhood does not exceed half of the distance between two anchor points.
In step S2, the GCNet structure is as shown in fig. 2, the GCNet includes two parts, namely an encoder part and a decoder part, and the specific goals of the encoder part are: extracting the coordinate characteristics of the input incomplete point cloud; the specific structure of the encoder part is as follows: firstly, inputting point cloud coordinates, obtaining a feature vector of the point cloud through a layer of convolution layer, then dividing a space where the point cloud is located into w multiplied by w anchor points, averaging vectors in grids corresponding to the anchor points so as to realize grid structurization of features, and finally extracting high-dimensional features through 3D convolution and residual connection; the specific goals of the decoder part are: generating point cloud coordinates of the missing part through the high-dimensional features extracted by the decoder part; the specific structure of the decoder part is as follows: the input high-dimensional features are subjected to 3D transposition convolution to realize feature up-sampling, and the feature up-sampling is used for predicting the positive and negative properties of anchor points and the coordinate information of the diffused point cloud in the neighborhood of the positive anchor points to generate the final complete point cloud.
In step S3, the loss function includes BCELoss and CD, and the specific formula of BCELoss is:
wherein XiTo predict value, yiFor labels, the specific formula for a CD is:
wherein CD (S)1,S2) Point cloud set S representing predictive generation1And a real point cloud set S2The Euclidean space average closest distance between; and x and y respectively represent one point cloud in the point cloud set and consist of three coordinates of x, y and z.
The embodiment 1-3 is used for visually displaying the completion effect of the GCNet. And after the GCNet model is trained in the training set.
Example 1:
and selecting an incomplete aircraft point cloud in the test set as an input, and outputting an anchor point and the complete aircraft point cloud. The completion effect is shown in fig. 3.
Example 2:
and selecting an incomplete aircraft point cloud in the test set as an input, and outputting an anchor point and the complete aircraft point cloud. The completion effect is shown in fig. 4.
Example 3:
and selecting an incomplete aircraft point cloud in the test set as an input, and outputting an anchor point and the complete aircraft point cloud. The completion effect is shown in fig. 5.
The foregoing merely represents preferred embodiments of the invention, which are described in some detail and detail, and therefore should not be construed as limiting the scope of the invention. It should be noted that, for those skilled in the art, various changes, modifications and substitutions can be made without departing from the spirit of the present invention, and these are all within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.
Claims (8)
1. A point cloud completion method based on anchor point detection is characterized in that: the method comprises the following steps:
s1, data set acquisition and data preprocessing: acquiring a point cloud public data set ShapeNet, and deleting local point clouds in complete point cloud data by using local deletion operation so as to construct incomplete point clouds;
s2, constructing a neural network model: based on the point cloud completion thought from rough to fine, firstly, judging through the positive and negative properties of an anchor point to obtain sparse point cloud with regular shape and uniform density, then obtaining dense point cloud through the diffusion of the sparse point cloud, and constructing a GCNet network model for realizing the thought;
s3, model training and model saving: and (3) training the network by using an Adam optimizer, reducing a loss function, improving a point cloud completion effect, and storing the model when the loss function of the model tends to be stable.
2. The point cloud complementing method based on anchor point detection according to claim 1, wherein: in step S1, the public data set sharenet includes 16 kinds of point cloud data, each individual point cloud data file (.pts) of all kinds is composed of n point cloud coordinates x, y, z in a three-dimensional coordinate system, and the value of n is different.
3. The point cloud complementing method based on anchor point detection according to claim 1, wherein: in step S1, the specific method of the local deletion operation includes definition and selection of a key point and deletion of a point cloud adjacent to the key point.
4. The point cloud complementing method based on anchor point detection as claimed in claim 3, wherein: in step S1, the specific method for defining and selecting the key points includes: taking 4 points of (1,0,0), (0,0,1), (1,0,1) and (-1,0,0) as key points, wherein three numbers in (b) represent coordinates of x, y and z in a three-dimensional coordinate system; in order to simulate incomplete point clouds with different parts missing, 4 key points are sequentially used as central points, and point cloud data near the central point are deleted; the method comprises the following specific steps of: and calculating the distances between all points in the point cloud and the center, and deleting 512 point clouds closest to the center point.
5. The point cloud complementing method based on anchor point detection according to claim 1, wherein: in the step S2, in the sparse point cloud with regular shape and uniform density obtained through the positive and negative judgment of the anchor points, only three times of the number of positive samples are selected from the negative samples for training; in addition, in the selected negative samples, the negative samples are further divided into the negative samples which are difficult to train and the negative samples which are easy to train by setting a threshold value d, and the negative samples which are difficult to train and have the distance smaller than the threshold value d are difficult to train; the number ratio of the negative samples which are difficult to train to the negative samples which are easy to train is 2: 1.
6. the point cloud complementing method based on anchor point detection according to claim 1, wherein: in the step S2, in the dense point cloud obtained by the diffusion of the sparse point cloud, point cloud coordinates in the neighborhood of each anchor point are generated through a neural network on the basis of the sparse point cloud (sample of the anchor point), and the maximum range of the neighborhood is limited by tanh so that the maximum range does not exceed half of the distance between two anchor points.
7. The point cloud complementing method based on anchor point detection according to claim 1, wherein: in step S2, the GCNet includes two parts, namely an encoder part and a decoder part, and the specific goals of the encoder part are: extracting the coordinate characteristics of the input incomplete point cloud; the specific structure of the encoder part is as follows: firstly, inputting point cloud coordinates, obtaining a feature vector of the point cloud through a layer of convolution layer, then dividing a space where the point cloud is located into w multiplied by w anchor points, averaging vectors in grids corresponding to the anchor points so as to realize grid structurization of features, and finally extracting high-dimensional features through 3D convolution and residual connection; the specific goals of the decoder part are: generating point cloud coordinates of the missing part through the high-dimensional features extracted by the decoder part; the specific structure of the decoder part is as follows: the input high-dimensional features are subjected to 3D transposition convolution to realize feature up-sampling, and the feature up-sampling is used for predicting the positive and negative properties of anchor points and the coordinate information of the diffused point cloud in the neighborhood of the positive anchor points to generate the final complete point cloud.
8. The point cloud complementing method based on anchor point detection according to claim 1, wherein: in step S3, the loss function includes BCELoss and CD, and the specific formula of BCELoss is:
wherein XiTo predict value, yiFor labels, the specific formula for a CD is:
wherein CD (S)1,S2) Point cloud set S representing predictive generation1And a real point cloud set S2The Euclidean space average closest distance between; and x and y respectively represent one point cloud in the point cloud set and consist of three coordinates of x, y and z.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111504436.0A CN114359510A (en) | 2021-12-10 | 2021-12-10 | Point cloud completion method based on anchor point detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111504436.0A CN114359510A (en) | 2021-12-10 | 2021-12-10 | Point cloud completion method based on anchor point detection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114359510A true CN114359510A (en) | 2022-04-15 |
Family
ID=81099930
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111504436.0A Pending CN114359510A (en) | 2021-12-10 | 2021-12-10 | Point cloud completion method based on anchor point detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114359510A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117173650A (en) * | 2023-11-02 | 2023-12-05 | 浙江华是科技股份有限公司 | Ship measurement and identification method and system based on laser radar |
-
2021
- 2021-12-10 CN CN202111504436.0A patent/CN114359510A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117173650A (en) * | 2023-11-02 | 2023-12-05 | 浙江华是科技股份有限公司 | Ship measurement and identification method and system based on laser radar |
| CN117173650B (en) * | 2023-11-02 | 2024-01-26 | 浙江华是科技股份有限公司 | Ship measurement and identification method and system based on laser radar |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105512289B (en) | Image search method based on deep learning and Hash | |
| CN107767453B (en) | Building LIDAR point cloud reconstruction optimization method based on rule constraint | |
| CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
| CN104954185B (en) | A kind of cloud computing load predicting method based on depth confidence net | |
| CN106295506A (en) | A kind of age recognition methods based on integrated convolutional neural networks | |
| CN104268593A (en) | Multiple-sparse-representation face recognition method for solving small sample size problem | |
| CN108121950B (en) | Large-pose face alignment method and system based on 3D model | |
| CN112085072B (en) | Cross-modal retrieval method of sketch retrieval three-dimensional model based on space-time characteristic information | |
| CN109711283A (en) | A kind of joint doubledictionary and error matrix block Expression Recognition algorithm | |
| CN108879732A (en) | Transient stability evaluation in power system method and device | |
| CN110991553A (en) | BIM model comparison method | |
| WO2022000522A1 (en) | Saliency characteristics-based simulation incomplete point cloud mask generation method | |
| CN110765788A (en) | Knowledge graph embedding method based on implicit translation model | |
| CN109754122A (en) | A kind of Numerical Predicting Method of the BP neural network based on random forest feature extraction | |
| CN109389553A (en) | Meteorological causes isopleth interpolation method based on T batten | |
| CN114359510A (en) | Point cloud completion method based on anchor point detection | |
| CN110659702A (en) | Calligraphy copybook evaluation system and method based on generative confrontation network model | |
| CN114972323A (en) | Defect picture generation method and system for model training | |
| CN113256543A (en) | Point cloud completion method based on graph convolution neural network model | |
| CN110363848A (en) | A kind of method for visualizing and device of the pore network model based on digital cores | |
| CN115953330B (en) | Texture optimization method, device, equipment and storage medium for virtual scene image | |
| CN111353525A (en) | Modeling and missing value filling method for unbalanced incomplete data set | |
| CN104200222B (en) | Object identifying method in a kind of picture based on factor graph model | |
| CN114155560B (en) | Light weight method of high-resolution human body posture estimation model based on space dimension reduction | |
| CN114861917A (en) | Knowledge graph inference model, system and inference method for Bayesian small sample learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |