CN114092580A - Three-dimensional point cloud data compression method and system based on deep learning - Google Patents
Three-dimensional point cloud data compression method and system based on deep learning Download PDFInfo
- Publication number
- CN114092580A CN114092580A CN202111294880.4A CN202111294880A CN114092580A CN 114092580 A CN114092580 A CN 114092580A CN 202111294880 A CN202111294880 A CN 202111294880A CN 114092580 A CN114092580 A CN 114092580A
- Authority
- CN
- China
- Prior art keywords
- point cloud
- cloud data
- semantic
- model
- octree
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/40—Tree coding, e.g. quadtree, octree
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/002—Image coding using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention provides a three-dimensional point cloud data compression method and system based on deep learning, and relates to the technical field of point cloud data compression processing, wherein the three-dimensional point cloud data compression method based on deep learning comprises the following steps: obtaining semantic tags of a point cloud data model based on PointNet neural network, adopting a Mean-Shift algorithm to perform component segmentation on the point cloud data model based on the semantic tags, optimally aligning a single item obtained by separation with a standard model based on semantics, organizing and sequencing, and then adopting an octree algorithm to realize data compression. The proposed method improves the coding performance of octrees by semantic labeling. Compared with the prior art, the method can obtain lower mean square error under the same octree algorithm rate, thereby achieving better compression effect.
Description
Technical Field
The invention relates to the technical field of point cloud data compression processing, in particular to a three-dimensional point cloud data compression method and system based on deep learning.
Background
The point cloud model segmentation is the key of robust target identification and has important significance for autonomous and automatic driving of the robot and the like. At present, Maturana et al propose an efficient point cloud segmentation architecture that enables 3D object recognition by occupying a volume with a mesh. In contrast, due to unnecessary redundancy, Qi et al propose a unified framework that locally processes each point to improve the efficiency of shape segmentation and to be robust to alignment invariance of input points. Landrieu et al use a hyper-point map, which is an element that segments a large-scale point cloud scene into geometrically uniform shapes to represent the three-dimensional shape segmentation. Brostow et al propose an algorithm for semantic segmentation of three-dimensional shapes by projecting 5 3D indices back to a two-dimensional image plane to model the spatial layout and environment of a moving point cloud.
Most of the existing three-dimensional grid compression methods are based on geometric structures, and due to the irregular format, most researchers convert the data into regular three-dimensional voxel grids or images, so that weight sharing, kernel parameter optimization and the like can be conveniently carried out through convolution operation in deep learning. However, this can make the data unnecessarily large and cause problems.
At present, many different technologies, such as wavelet transformation, graph transformation and hierarchical transformation, are available for the compression research of three-dimensional point cloud models. In this field, researchers have also developed efficient and effective 3D animation compression methods. These methods improve compression efficiency compared to static shapes by studying temporal redundancy of motion and spatial redundancy of surfaces.
In summary, the current technical scheme can realize point cloud compression, but the compression efficiency is still the direction of the current domestic research.
Chinese patent CN1110135227A entitled "automatic segmentation method of laser point cloud outdoor scene based on machine learning" discloses a method for iteratively compressing data by calculating feature vectors based on three-dimensional convolutional neural network and combining cost function. However, semantic segmentation is carried out based on the PointNet neural network, and the point cloud data is compressed by utilizing the octree algorithm.
Disclosure of Invention
Aiming at the specific problem of low compression efficiency in the background technology, the three-dimensional point cloud data compression method and system based on deep learning are provided, the optimal alignment with a standard model based on semantics is calculated by combining a three-dimensional model with semantic marks, and then an octree algorithm is used for compression, so that the octree coding performance is superior, and the compression efficiency is improved.
In order to achieve the above object, the present invention adopts the following aspects.
A three-dimensional point cloud data compression method based on deep learning comprises the following steps:
101, performing semantic segmentation on an input point cloud data model through a PointNet neural network to obtain a semantic combination model based on semantic labels;
102, carrying out component segmentation based on the semantic combination model to obtain a single item;
103, calculating the optimal alignment of the point cloud data and a standard model based on semantics based on the single item, and organizing and sequencing the point cloud data to obtain an ordered point cloud data model;
and 104, acquiring the spatial characteristics of the point cloud data model by using an octree representation method based on the ordered point cloud data model, calculating three-dimensional object classification by combining the spatial characteristics, and performing fluidization treatment on items in each class to realize compression.
Preferably, in the deep learning-based three-dimensional point cloud data compression method, step 102 specifically includes: and carrying out component segmentation on the semantic combination model by adopting a density-based nonparametric clustering algorithm so as to separate the semantic combination model into single items.
Preferably, in the three-dimensional point cloud data compression method based on deep learning, the non-parametric clustering algorithm based on density adopts a Mean-Shift algorithm.
Preferably, in the deep learning-based three-dimensional point cloud data compression method, step 103 further includes aligning the single item with a semantic-based standard model by using an iterative closest point algorithm.
Preferably, in the deep learning-based three-dimensional point cloud data compression method, the semantic-based standard model is used for automatically selecting objects with a medium number of points from each semantic set
Preferably, in the method for compressing three-dimensional point cloud data based on deep learning, step 103 specifically includes: and organizing and sequencing the point cloud data by adopting a KD-tree algorithm.
Preferably, in the deep learning-based three-dimensional point cloud data compression method, step 104 specifically includes: from the spatial feature representation, a classification of all individual items is computed using mixed-layer evaluation, and the items within each class are then fluidized based on the classification.
Preferably, in the deep learning-based three-dimensional point cloud data compression method, step 104 specifically includes: and performing three-dimensional object classification on the point cloud data model by adopting a K-Means clustering algorithm to realize grouping of point cloud data with similar octree structure space characteristics.
A three-dimensional point cloud data compression system based on deep learning comprises:
the semantic segmentation module is used for performing semantic segmentation on the input point cloud data model based on a PointNet neural network to obtain a semantic combination model based on semantic labels;
the component segmentation module is used for carrying out component segmentation on the basis of the semantic combination model to obtain a single item;
the optimal alignment module is used for calculating optimal alignment between the point cloud data and a standard model based on semantics based on the single item, and then organizing and sequencing the point cloud data to obtain an ordered point cloud data model;
and the octree coding module is used for acquiring the spatial characteristics of the point cloud data model by utilizing an octree representation method based on the ordered point cloud data model, calculating three-dimensional object classification by combining the spatial characteristics, and performing fluidization treatment on items in each class to realize compression.
In summary, due to the adoption of the technical scheme, the invention at least has the following beneficial effects:
the invention provides a three-dimensional point cloud data compression method and system based on deep learning. Secondly, decomposing the semantic combination model into single items, calculating the optimal alignment with a standard object, obtaining spatial features by using an octree representation method, calculating classification based on the spatial features, and finally completing stream processing on the items in each class to realize compression. The three-dimensional point cloud data compression method and the system improve the encoding performance of the octree by utilizing semantic classification, further reduce the data redundancy and further improve the compression efficiency.
Drawings
FIG. 1 is a flow chart of a large scale three-dimensional point cloud scene flow modeling based on semantic classification according to an exemplary embodiment of the present invention;
FIG. 2 is a diagram of a semantic composition model based on "chair" labels according to an exemplary embodiment of the present invention;
FIG. 3 is a diagram of two individual item maps based on "chair" tags and a semantic-based standard model according to an exemplary embodiment of the present invention;
FIG. 4 is a classification flow graph based on a hierarchical octree structure according to an exemplary embodiment of the present invention;
FIG. 5 is a graph comparing rate-distortion curves based on the presence or absence of semantic classification according to an exemplary embodiment of the present invention;
FIG. 6 is a flow diagram of a large scale point cloud model with 41,353,055 points at different levels of detail according to an exemplary embodiment of the invention;
FIG. 7 is a progressive flow graph of chair (top) and Room2 (bottom) data using the proposed model, according to an exemplary embodiment of the present invention;
FIG. 8 is a compression system based on a PointNet neural network model in accordance with an exemplary embodiment of the present invention;
the labels in the figure are: 100-semantic segmentation module, 200-component segmentation module, 300-optimal alignment module, and 400-octree coding module.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings and embodiments, so that the objects, technical solutions and advantages of the present invention will be more clearly understood. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Example 1
The invention relates to a three-dimensional point cloud data compression method based on deep learning, wherein a large-scale three-dimensional point cloud scene flow model based on semantic classification is shown in figure 1, and the method comprises the following steps:
specifically, a classical depth network-based technology is applied to calculate semantic labels of an input large-scale three-dimensional model, and the large-scale three-dimensional model is set as M, wherein M = { Pi | i = 1,2, ·, n }, and n is a point number. Semantic labeling is performed on each vertex Pi, which can be expressed as:
where t is the number of semantic classes.
Using semantic tags, we obtain a semantic combination model containing a set of items with similar topology, and further obtain the semantic combination model by performing semantic segmentation based on "chair" tags on point cloud data models, where the point cloud data models all have "chair" topology but have differences in color, size, direction, etc., as shown in fig. 2.
And 102, carrying out component segmentation based on the semantic combination model to obtain a single item.
Specifically, redundancy is studied by further separating the semantic composition model. The method adopts the Mean-Shift algorithm as a non-parametric clustering algorithm based on density to carry out segmentation so as to separate and obtain a single item based on semantic mark. And performing component segmentation on the semantic combination model with the chair mark to obtain a single model based on the chair mark, namely a single chair shown in fig. 3, namely a single chair 1, a single chair 2 and a standard chair model from left to right.
The semantic composition model is divided into S individual items, which can be expressed as:
wherein Mj is a semantic set; cj is the number of categories, which is influenced by the maximum distance r between the point of computation and the current center.
And 103, calculating the optimal alignment of the point cloud data and the standard model based on the semantics based on the single item, and organizing the point cloud data to obtain an ordered point cloud data model.
In particular, since there is a difference between single items with the same semantics, it is not beneficial to reduce the spatial redundancy. To reduce spatial redundancy, a single item under semantics is aligned with a standard semantic-based model. As shown in fig. 3, the optimal alignment of the individual chair with the standard chair model is to align the individual chairs 1,2 with similar color, size and orientation to the standard chair model with the color, size and orientation of the standard chair model as the alignment condition.
The invention adopts ICP algorithm, which comprises the following steps:
under a single item of semantic identity, the alignment matching transformation of a first set of points, a second set of points representing space, minimizes the objective function of:
wherein, the target point set is a single item; is a reference point set, namely a standard model based on semantics;Rin order to be the rotation factor,Tis the translation factor.
The essence of the ICP algorithm is an optimal matching algorithm based on the least square method, which repeats the process of 'determining a corresponding relation point set-calculating an optimal rigid body transformation' in order to find the rotation between a target point set and a reference point setRAnd translationTAnd transforming, wherein the rigid transformation is processed by a quaternion method. It will be appreciated that by computing a rigid transformation matrix, the closest point between a single item and the semantic-based standard model is iteratively searched, while the computation returns a match error. This process is repeated until some convergence condition is satisfied indicating a correct match. In the next step, no parallax operation is performed for all models.
Preferably, the semantic-based standard model is an object with a moderate number of points automatically selected from each semantic set.
Specifically, due to the fact that three-dimensional point cloud data are large, the point cloud data are organized and sorted based on a KD-tree (K-dimensional tree), and the KD-tree is a binary search tree with other constraint conditions and used for organizing and representing a point set in a K-dimensional space. Its nodes are represented by pairs of (a, V) values, where a is the current partition dimension and V is the partition value. In a K-dimensional data space, (a, V) forms a K-1-dimensional hyperplane, the hyperplane is utilized to divide the data space, and the multidimensional space is divided into a subspace with an a value smaller than V and a subspace with an a value larger than or equal to V, wherein the a adopts the dimension with the largest variance among the K dimensions.
The KD-tree may be used to build indices into multidimensional spatial data sets or data blocks. When the KD tree is used for data query, only one dimension needs to be compared for the condition judgment of each step of node, so that the neighborhood of a certain data point can be quickly searched by alternately comparing the attribute values of different dimensions without knowing any topological relation among data, and an ordered point cloud data model is obtained to optimize the coding efficiency of the octree.
And 104, acquiring the spatial characteristics of the point cloud data model by using an octree representation method based on the ordered point cloud data model, calculating three-dimensional object classification by combining the spatial characteristics, and performing fluidization treatment on items in each class to realize compression.
Specifically, using octree to represent point cloud data model, the input point cloud model can be represented by recording node attributes of 8 child nodes of a non-leaf node, wherein each non-leaf node is converted into 8-bit binary data, resulting in semantic classification based on octree hierarchy, as shown in fig. 4. And then, carrying out hierarchical traversal on the octree to obtain a one-dimensional vector:
wherein l is a spreading layer. Octree coding can be extended to different levels by determining specific levels of octrees. Thus, given a particular level, each node represents a point within all boxes, as shown in Table 1, with time units of seconds.
TABLE 1
Specifically, based on the octree hierarchical structure point cloud data model, spatial feature classification of a single item is calculated, and different mixed classification items based on semantic labels of the single item are obtained. In order to improve the compression effect, the maximum spatial redundancy among all mixed classification items is further explored:
wherein, W represents the weight of each layer of the octree, the weight of the bottom layer is 1, and the weight of each layer of the upper layer is multiplied by 8. The output of f2 is the mixed tier value of each individual item.
Wherein the K-Means clustering algorithm is utilized to group the single items with similar octree space structures. The K-Means clustering algorithm is that on the basis of density parameters, the Euclidean distance is utilized to calculate the density parameters of a data model, after all the density parameters are obtained, K centers are obtained, if the data objects A have the same distance to the K centers, the density distances from the data objects in the cluster to the data objects A are sorted at the moment, the minimum density distance is selected, and the data objects A are classified into the corresponding classes, so that the data objects and the adjacent data objects can be more compact, and the data objects with similar octree space structures can be classified into the same group.
Finally, the number of layers of octree can be gradually increased in the data stream, so that the point cloud data can be displayed more finely, and therefore, a large number of details of the outline of the point cloud data model can be observed, as shown in fig. 6 and 7, by increasing the number of layers of octree, a clearer data model based on color and outline from left to right is obtained.
Meanwhile, the influence of semantic segmentation on the coding performance represented by the octree can be seen from experimental data, the semantic segmentation obviously improves the coding performance of the octree, and as shown in fig. 5, the mean square error of the data which is not optimally aligned is higher than that of the data after alignment in a proper compression ratio range and under the condition of the same compression ratio. That is, at the same rate, we can get a lower mean square error because the semantic segmentation has further explored the redundancy in the various items in the point cloud model.
The invention also provides a three-dimensional point cloud data compression system based on deep learning. Fig. 8 schematically shows a structure diagram of a three-dimensional point cloud data compression system based on deep learning, which specifically includes:
the semantic segmentation module is used for performing semantic segmentation on the input point cloud data model based on a PointNet neural network to obtain a semantic combination model based on semantic labels;
the component segmentation module is used for carrying out component segmentation on the basis of the semantic combination model to obtain a single item;
the optimal alignment module is used for calculating optimal alignment between the point cloud data and a standard model based on semantics based on the single item, and then organizing and sequencing the point cloud data to obtain an ordered point cloud data model;
and the octree coding module is used for acquiring the spatial characteristics of the point cloud data model by utilizing an octree representation method based on the ordered point cloud data model, calculating three-dimensional object classification by combining the spatial characteristics, and performing fluidization treatment on items in each class to realize compression.
The semantic segmentation module 100 performs semantic segmentation on the input point cloud data model based on a PointNet neural network, and outputs the data model with semantic labels to the component segmentation module 200; the component segmentation module 200 performs semantic tag-based component segmentation on point cloud data and outputs a plurality of semantic-based single items; the optimal alignment module 300 acts on spatial redundancy caused by differences of the single items, performs preliminary spatial sorting on the point cloud data model and outputs an ordered point cloud data model; the octree coding module 400 performs hierarchical traversal based on the octree space structure on the ordered point cloud data model, and finally obtains scene codes through octree structure classification.
Claims (9)
1. A three-dimensional point cloud data compression method based on deep learning is characterized by comprising the following steps:
101, performing semantic segmentation on an input point cloud data model through a PointNet neural network to obtain a semantic combination model based on semantic labels;
102, carrying out component segmentation based on the semantic combination model to obtain a single item;
103, calculating the optimal alignment of the point cloud data and a standard model based on semantics based on the single item, and organizing and sequencing the point cloud data to obtain an ordered point cloud data model;
and 104, acquiring the spatial characteristics of the point cloud data model by using an octree representation method based on the ordered point cloud data model, calculating three-dimensional object classification by combining the spatial characteristics, and performing fluidization treatment on items in each class to realize compression.
2. The method according to claim 1, wherein the step 102 further comprises: and carrying out component segmentation on the semantic combination model by adopting a density-based nonparametric clustering algorithm so as to separate the semantic combination model into single items.
3. The method of claim 2, wherein the density-based non-parametric clustering algorithm employs a Mean-Shift algorithm.
4. The method according to claim 1, wherein the step 103 further comprises: and aligning the single item with the standard model based on the semanteme by adopting an iterative closest point algorithm.
5. The method of claim 4, wherein the semantic-based criteria model is an automatic selection of objects with a medium number of points from each semantic set.
6. The method according to claim 1, wherein the step 103 further comprises: and organizing and sequencing the point cloud data by adopting a KD-tree algorithm.
7. The method according to claim 1, wherein the step 104 further comprises: from the spatial feature representation, a classification of all individual items is computed with mixed-layer evaluation.
8. The method according to claim 1, wherein the step 104 further comprises: and performing three-dimensional object classification on the point cloud data model by adopting a K-Means clustering algorithm to realize grouping of point cloud data with similar octree structure space characteristics.
9. A three-dimensional point cloud data compression system based on deep learning is characterized by comprising:
the semantic segmentation module is used for performing semantic segmentation on the input point cloud data model based on a PointNet neural network to obtain a semantic combination model based on semantic labels;
the component segmentation module is used for carrying out component segmentation on the basis of the semantic combination model to obtain a single item;
the optimal alignment module is used for calculating optimal alignment between the point cloud data and a standard model based on semantics based on the single item, and then organizing and sequencing the point cloud data to obtain an ordered point cloud data model;
and the octree coding module is used for acquiring the spatial characteristics of the point cloud data model by utilizing an octree representation method based on the ordered point cloud data model, calculating three-dimensional object classification by combining the spatial characteristics, and performing fluidization treatment on items in each class to realize compression.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111294880.4A CN114092580B (en) | 2021-11-03 | 2021-11-03 | Three-dimensional point cloud data compression method and system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111294880.4A CN114092580B (en) | 2021-11-03 | 2021-11-03 | Three-dimensional point cloud data compression method and system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114092580A true CN114092580A (en) | 2022-02-25 |
| CN114092580B CN114092580B (en) | 2022-10-21 |
Family
ID=80298807
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111294880.4A Active CN114092580B (en) | 2021-11-03 | 2021-11-03 | Three-dimensional point cloud data compression method and system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114092580B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115761023A (en) * | 2022-12-02 | 2023-03-07 | 同济大学 | Three-dimensional point cloud compression system and method based on point cloud matrix singular value characteristics |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109117718A (en) * | 2018-07-02 | 2019-01-01 | 东南大学 | A kind of semantic map structuring of three-dimensional towards road scene and storage method |
| CN111754637A (en) * | 2020-06-30 | 2020-10-09 | 华东交通大学 | Large-scale three-dimensional face synthesis system with suppressed sample similarity |
| US20210019918A1 (en) * | 2019-07-15 | 2021-01-21 | Peking Universtiy Shenzhen Graduate School | Point cloud geometric compression method based on depth auto-encoder |
| CN112287939A (en) * | 2020-10-29 | 2021-01-29 | 平安科技（深圳）有限公司 | Three-dimensional point cloud semantic segmentation method, device, equipment and medium |
| CN112927234A (en) * | 2021-02-25 | 2021-06-08 | 中国工商银行股份有限公司 | Point cloud semantic segmentation method and device, electronic equipment and readable storage medium |
| CN113449736A (en) * | 2021-01-14 | 2021-09-28 | 浙江工业大学 | Photogrammetry point cloud semantic segmentation method based on deep learning |
| CN113486902A (en) * | 2021-06-29 | 2021-10-08 | 南京航空航天大学 | Three-dimensional point cloud classification algorithm automatic selection method based on meta-learning |
-
2021
- 2021-11-03 CN CN202111294880.4A patent/CN114092580B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109117718A (en) * | 2018-07-02 | 2019-01-01 | 东南大学 | A kind of semantic map structuring of three-dimensional towards road scene and storage method |
| US20210019918A1 (en) * | 2019-07-15 | 2021-01-21 | Peking Universtiy Shenzhen Graduate School | Point cloud geometric compression method based on depth auto-encoder |
| CN111754637A (en) * | 2020-06-30 | 2020-10-09 | 华东交通大学 | Large-scale three-dimensional face synthesis system with suppressed sample similarity |
| CN112287939A (en) * | 2020-10-29 | 2021-01-29 | 平安科技（深圳）有限公司 | Three-dimensional point cloud semantic segmentation method, device, equipment and medium |
| CN113449736A (en) * | 2021-01-14 | 2021-09-28 | 浙江工业大学 | Photogrammetry point cloud semantic segmentation method based on deep learning |
| CN112927234A (en) * | 2021-02-25 | 2021-06-08 | 中国工商银行股份有限公司 | Point cloud semantic segmentation method and device, electronic equipment and readable storage medium |
| CN113486902A (en) * | 2021-06-29 | 2021-10-08 | 南京航空航天大学 | Three-dimensional point cloud classification algorithm automatic selection method based on meta-learning |
Non-Patent Citations (6)
| Title |
|---|
| GUOLIANG LUO等: "Dynamic data reshaping for 3D mesh animation compression", 《MULTIMEDIA TOOLS AND APPLICATIONS》 * |
| WEI ZENG等: "3DContextNet: K-d Tree Guided Hierarchical Learning of Point Clouds Using Local and Global Contextual Cues", 《ECCV 2018》 * |
| 朱威等: "基于动态图卷积和空间金字塔池化的点云深度学习网络", 《计算机科学》 * |
| 杨刚等: "基于AQPSO的RBF神经网络自组织学习", 《控制与决策》 * |
| 杨晓文等: "基于八叉树的卷积神经网络三维模型分割", 《计算机工程与设计》 * |
| 顾军华等: "基于点云数据的分割方法综述", 《燕山大学学报》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115761023A (en) * | 2022-12-02 | 2023-03-07 | 同济大学 | Three-dimensional point cloud compression system and method based on point cloud matrix singular value characteristics |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114092580B (en) | 2022-10-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Guo et al. | Deep learning for 3d point clouds: A survey | |
| Rao et al. | Spherical fractal convolutional neural networks for point cloud recognition | |
| Ahmed et al. | Deep learning advances on different 3D data representations: A survey | |
| Yang et al. | Content-based 3-D model retrieval: A survey | |
| Zhang et al. | Cascaded context pyramid for full-resolution 3d semantic scene completion | |
| US8429174B2 (en) | Methods, systems, and data structures for performing searches on three dimensional objects | |
| CN107085731B (en) | Image classification method based on RGB-D fusion features and sparse coding | |
| Hu et al. | RGB-D semantic segmentation: a review | |
| Bobkov et al. | Noise-resistant deep learning for object classification in three-dimensional point clouds using a point pair descriptor | |
| Bewley et al. | Advantages of exploiting projection structure for segmenting dense 3D point clouds | |
| CN113515656A (en) | Multi-view target identification and retrieval method and device based on incremental learning | |
| CN114092580B (en) | Three-dimensional point cloud data compression method and system based on deep learning | |
| CN113628329B (en) | Zero-sample sketch three-dimensional point cloud retrieval method | |
| Akgül et al. | Density-based 3D shape descriptors | |
| ElNaghy et al. | Taxonomy for 3D content-based object retrieval methods | |
| Xiao et al. | Enhanced 3-D modeling for landmark image classification | |
| CN111597367B (en) | Three-dimensional model retrieval method based on view and hash algorithm | |
| CN112200248A (en) | Point cloud semantic segmentation method, system and storage medium under urban road environment based on DBSCAN clustering | |
| Wei et al. | Representation and retrieval of 3D CAD models in parts library | |
| CN109272013B (en) | Similarity measurement method based on learning | |
| CN111414802B (en) | Protein data characteristic extraction method | |
| Li et al. | Semi-supervised implicit scene completion from sparse LiDAR | |
| Dimou et al. | Fast 3D scene segmentation and partial object retrieval using local geometric surface features | |
| Yang et al. | Unsupervised co-segmentation of 3d shapes based on components | |
| Wang et al. | Geodesic-HOF: 3D reconstruction without cutting corners |
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 |