CN108319957A - A kind of large-scale point cloud semantic segmentation method based on overtrick figure - Google Patents
A kind of large-scale point cloud semantic segmentation method based on overtrick figure Download PDFInfo
- Publication number
- CN108319957A CN108319957A CN201810132821.9A CN201810132821A CN108319957A CN 108319957 A CN108319957 A CN 108319957A CN 201810132821 A CN201810132821 A CN 201810132821A CN 108319957 A CN108319957 A CN 108319957A
- Authority
- CN
- China
- Prior art keywords
- overtrick
- point
- input
- cloud
- point cloud
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/2163—Partitioning the feature space
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Abstract
The present invention proposes a kind of large-scale point cloud semantic segmentation method based on overtrick figure, and main contents include：Geometry uniform segmentation, overtrick figure is built, embedded overtrick, semantic segmentation, training and test, its process is, a cloud is first divided into geometry, referred to as overtrick, cloud will be entirely put as input using this unsupervised step, overtrick figure is calculated in geometric zoning, then the dimension of fixed size is selected in each overtrick, descriptor is calculated by embedded vector, finally since the figure of overtrick is smaller than the figure established on original point cloud, pass through the deep learning algorithm accumulated based on picture scroll, classified to its node using abundant edge feature, overtrick is embedded according to the information refinement that extreme edges is transmitted.The present invention solves the problems, such as that the semantic segmentation on large-scale three dimensional point cloud, overtrick figure handle large-scale data on the basis of deep learning frame, and segmentation efficiency is improved while retaining minor detail.
Description
Technical field
The present invention relates to semantic segmentation fields, more particularly, to a kind of large-scale point cloud semantic segmentation based on overtrick figure
Method.
Background technology
Semantic segmentation is exactly that machine is divided and identifies the content in image automatically, it may be said that is the basic of image understanding
Technology is held the balance in automated driving system, unmanned plane application and Wearable application.It is well known that image be by
Many pixels composition, and " semantic segmentation " as the term suggests be exactly to be divided pixel according to the difference for expressing semantic meaning in image
It cuts, semantic segmentation is an important branch in artificial intelligence field, is in machine vision technique about the important of image understanding
One ring detects the obstacles figure such as pedestrian, vehicle or trees and electric pole in automatic Pilot technology in recent years in vehicle-mounted camera
As after, background computer can divide the image into classification automatically, and driver is prompted to take corresponding measures to keep clear.In addition, semantic point
Medical image can be handled pixel-by-pixel by cutting, by the image that certain semantic segmentation Medical Instruments takes, can detect heart disease,
Divide tumour, saprodontia etc., to assist the diagnosis state of an illness.Further, it is also possible to install camera on unmanned plane, unmanned plane passes through
Surrounding enviroment are shot, building, plant, the road etc. in environment are split using semantic segmentation technology, to judge
Lu Dian.In robot application field, after robot receives instruction, built-in computer starts to call camera shooting periphery object simultaneously
Object is identified using image Segmentation Technology, can effectively get around barrier, is reached the instruction destination and is completed task, greatly
Ground facilitates people’s lives.Although the research in terms of cloud semantic segmentation is a lot of, be available data scale it is smaller and
Structural fuzzy, this leads to convolutional neural networks inefficiency when handling image on irregular data, therefore in large-scale three dimensional point
Semantic segmentation on cloud still remains challenge.
The present invention proposes a kind of large-scale point cloud semantic segmentation method based on overtrick figure, using based on deep learning
Frame handles the large-scale point cloud semantic segmentation of millions of points.A cloud is divided into geometry, referred to as overtrick first, is utilized
This unsupervised step will entirely put cloud as input, calculate overtrick figure in geometric zoning, then be selected in each overtrick
The dimension of fixed size calculates descriptor by embedded vector, finally since the figure of overtrick on original point cloud than establishing
Figure is small, by the deep learning algorithm accumulated based on picture scroll, is classified to its node using abundant edge feature, overtrick
The information refinement insertion transmitted according to extreme edges.The present invention solves the problems, such as the semantic segmentation on large-scale three dimensional point cloud, overtrick figure
Large-scale data is handled on the basis of deep learning frame, segmentation efficiency is improved while retaining minor detail.
Invention content
For data scale in semantic segmentation the problem of smaller and structural fuzzy, the purpose of the present invention is to provide a kind of bases
In the large-scale point cloud semantic segmentation method of overtrick figure, the extensive of millions of points is handled using the frame based on deep learning
Point cloud semantic segmentation.A cloud is first divided into geometry, referred to as overtrick, will entirely put cloud using this unsupervised step makees
For input, overtrick figure is calculated in geometric zoning, the dimension of fixed size is then selected in each overtrick, by embedded to gauge
It calculates descriptor and passes through the depth accumulated based on picture scroll finally since the figure of overtrick is smaller than the figure established on original point cloud
Algorithm is practised, is classified to its node using abundant edge feature, overtrick is embedded according to the information refinement that extreme edges is transmitted.
To solve the above problems, the present invention provides a kind of large-scale point cloud semantic segmentation method based on overtrick figure, master
The content is wanted to include：
(1) geometry uniform segmentation；
(2) overtrick figure is built；
(3) embedded overtrick；
(4) semantic segmentation；
(5) training and test.
Wherein, a cloud is divided into geometry, referred to as overtrick by the geometry uniform segmentation, unsupervised using this
The step of will entirely put cloud as input, calculate overtrick figure (SPG) in geometric zoning, each node of SPG corresponds to geometrically simple
The sub-fraction point cloud of single object, it is contemplated that be semantically it is uniform, parameter by dot cloud be down-sampled to hundreds of points come
It indicates.
Further, local geometric complexity can be divided and be adapted to the subregion, Universal Energy model, therefore utilize logical
With energy model come computational efficiency, input point cloud C is regarded as the point of one group of n three-dimensional point composition, it is defined by each point i ∈ C
The positions 3D pi, and other observed values are defined as o color or intensity etc.i, for each point, computational geometry feature dg,
For characterizing the shape of its local neighborhood, variable sampling density is compensated by adaptive neighborhood, using the linearity, flatness and
Three dimension values are scattered, and introduce verticality, calculate the elevation each put to indicate p after normalizationiZ on entirely input cloud
Axial coordinate, geometry uniform segmentation is the constant connection component of optimization problem reconciliation, given by following formula：
Wherein [≠ 0] isFunction, whenWhen, [≠ 0] is equal to 0, other feelings
It is equal to 1 under condition, coefficient μ is regularization intensity, and for determining the roughness of institute's scoring area, constant communication component S={ S1,…,
SkBe equation (1) solution, for defining geometrically simple element.
Wherein, overtrick figure structure, SPG is the structured representation of a cloud, defines a directional properties figureIts node is the set of overtrick S, and the syntople between overtrick is indicated with extreme edges ε, with one group of dfFeature is noted
Release extreme edges： Including the syntople between overtrick, by Gvor=(C, Evor) to be defined as complete input point cloud symmetrical
Voronoi adjacent maps, if EvorIn there are one edge, then S and T is two adjacent overtricks, and S and T are located at two
End：
EvorTwo overtricks are connected, the important of extreme edges (S, T) is obtained from the edge offset amount δ (S, T) of the two overtricks
Correlated characteristic：
δ (S, T)={ (pi-pj)|(i,j)∈Evor∩(S×T)} (3)
Extreme edges feature is exported by comparing the shapes and sizes of adjacent overtrick, is used | S | indicate the point for including in overtrick S
Number, λ1,λ2,λ3It indicates that each overtrick includes the position covariance of point, the characteristic length of shape is exported by covariance
(S)=λ1, surface (S)=λ1λ2And volume (S)=λ1λ2λ3, by successively decreasing, sequence sorts.
Wherein, the insertion overtrick, in each overtrick SiIn, select the dimension d of fixed sizez, pass through embedded vector zi
Descriptor is calculated, each overtrick is separately embedded；Select the spot net of a deep learning, in a network, input point
It is aligned by spatial alternation network, then by multilayer perceptron independent process, is finally summarized to indicate input shape first,
Middle input shape is a simple geometric object, indicates input shape using a small amount of point, and pass through a compact point net
Network executes insertion, is n by overtrick fast samplingp=128, to maintain effectively to calculate in batches and data to be promoted to increase, replace
Less than npOvertrick sampled, in principle from the point of view of this have no effect on the assessment in spot net maximum pond, however pass through and test table
It is bright：Less than or equal to nminp=40 overtrick can damage the overall performance of network in training, therefore set the insertion of overtrick to
Zero, make its semantic information that places one's entire reliance upon of classifying, in order to make spot net learn spatial distribution of different shapes, each overtrick is embedding
It is scaled to unit sphere before entering, using their normalization position p 'i, observation oiWith geometric properties fiTo indicate a little, to be
Coordination shape size, the raw metric diameter of overtrick is as the supplementary features after spot net maximum pond.
Wherein, the semantic segmentation is accumulated since the figure of overtrick is smaller than the figure established on original point cloud based on picture scroll
Deep learning algorithm can be classified to its node using abundant edge feature, to promote prolonged interaction to make
With.
Further, the segmentation, overtrick are embedded according to the information refinement that extreme edges is transmitted, specifically, each overtrick
SiThe hidden state being maintained in gating cycle unit (GRU) passes through embedded ziHidden state is initialized, is then used
Iteration t=1 ... T processing, in each iteration t, GRU is by its hidden stateWith an input informationAs defeated
Enter, and calculates its new hidden stateThe input information of overtrick iBy being used as hiding for adjacent overtrick j after calculating
StateWeighted sum, the practical weighting of extreme edges (j, i) depends on its attribute Fji, pass through the attribute of multilayer perceptron Θ
It calculates：
Wherein, ⊙ is Element-Level multiplication, and σ () is sigmoid function, and W. and b. are shared between all GRU train
Parameter, in equation (4)The update of expression standard GRU rules gates, in equation (5)It indicates to reset gate, in order to improve
Stability in training process, first by the input after linear transformation in equation (8)It is defined as ρ (a)：=(a-mean
(a))/(std (a)+∈) then converts the hidden state in equation (7)Wherein ∈ is a smaller constant.
Further, the gate is being inputted by equation (9)Gating information is hidden beforeGRU can be with
Input vector is reduced according to its hidden state, Θ returns a weight matrix to execute matrix-vector multiplication for each edge,
Although conventional convolution can be carried out by demonstrating it on grid, longer run time can be caused, occupy higher memory
And more parameters are generated, therefore a specific weight vector in edge is returned using equation (10), execute Element-Level multiplication.
Further, the hidden state, cascades hidden state, is connected in all time steps and hides shape
State, and linear transformation they generate parted pattern yi, given by following equation：
This makes in final classification, due to the increase of acceptance region, can utilize the dynamic of hidden state.
Wherein, the training and test, for training：It is embedded although the step of geometry is divided is unsupervised
Overtrick and semantic segmentation are by the way of intersecting entropy loss supervision, it is assumed that semantic nature is identical, and overtrick is being semantically uniform
, between the point that they are included, specify calibration label corresponding with most of labels, the Neighborhood Graph that match exponents is 3 carries out
Sampling, so that each SPG at most selections 512 are more than nminpOvertrick；For test：Complete mark SPG, in order to compensate for due to
Randomness caused by the sub-sampling of point cloud in spot net, the parted pattern obtained to 10 operations using different o'clock are put down
.
Description of the drawings
Fig. 1 is a kind of system framework figure of the large-scale point cloud semantic segmentation method based on overtrick figure of the present invention.
Fig. 2 is a kind of visible process figure of the large-scale point cloud semantic segmentation method based on overtrick figure of the present invention.
Fig. 3 is a kind of segmentation instance graph of the large-scale point cloud semantic segmentation method based on overtrick figure of the present invention.
Specific implementation mode
It should be noted that in the absence of conflict, the features in the embodiments and the embodiments of the present application can phase
It mutually combines, invention is further described in detail in the following with reference to the drawings and specific embodiments.
Fig. 1 is a kind of system framework figure of the large-scale point cloud semantic segmentation method based on overtrick figure of the present invention.Main packet
Include geometry uniform segmentation, overtrick figure structure, embedded overtrick, semantic segmentation, training and test.
Wherein, a cloud is divided into geometry, referred to as overtrick by the geometry uniform segmentation, unsupervised using this
The step of will entirely put cloud as input, calculate overtrick figure (SPG) in geometric zoning, each node of SPG corresponds to geometrically simple
The sub-fraction point cloud of single object, it is contemplated that be semantically it is uniform, parameter by dot cloud be down-sampled to hundreds of points come
It indicates.
Further, local geometric complexity can be divided and be adapted to the subregion, Universal Energy model, therefore utilize logical
With energy model come computational efficiency, input point cloud C is regarded as the point of one group of n three-dimensional point composition, it is defined by each point i ∈ C
The positions 3D pi, and other observed values are defined as o color or intensity etc.i, for each point, computational geometry feature dg,
For characterizing the shape of its local neighborhood, variable sampling density is compensated by adaptive neighborhood, using the linearity, flatness and
Three dimension values are scattered, and introduce verticality, calculate the elevation each put to indicate p after normalizationiZ on entirely input cloud
Axial coordinate, geometry uniform segmentation is the constant connection component of optimization problem reconciliation, given by following formula：
Wherein [≠ 0] isFunction, whenWhen, [≠ 0] is equal to 0, other feelings
It is equal to 1 under condition, coefficient μ is regularization intensity, and for determining the roughness of institute's scoring area, constant communication component S={ S1,…,
SkBe equation (1) solution, for defining geometrically simple element.
Wherein, overtrick figure structure, SPG is the structured representation of a cloud, defines a directional properties figureIts node is the set of overtrick S, and the syntople between overtrick is indicated with extreme edges ε, with one group of dfFeature is noted
Release extreme edges： Including the syntople between overtrick, by Gvor=(C, Evor) to be defined as complete input point cloud symmetrical
Voronoi adjacent maps, if EvorIn there are one edge, then S and T is two adjacent overtricks, and S and T are located at two
End：
EvorTwo overtricks are connected, the important of extreme edges (S, T) is obtained from the edge offset amount δ (S, T) of the two overtricks
Correlated characteristic：
δ (S, T)={ (pi-pj)|(i,j)∈Evor∩(S×T)} (3)
Extreme edges feature is exported by comparing the shapes and sizes of adjacent overtrick, is used | S | indicate the point for including in overtrick S
Number, λ1,λ2,λ3It indicates that each overtrick includes the position covariance of point, the characteristic length of shape is exported by covariance
(S)=λ1, surface (S)=λ1λ2And volume (S)=λ1λ2λ3, by successively decreasing, sequence sorts.
Wherein, the insertion overtrick, in each overtrick SiIn, select the dimension d of fixed sizez, pass through embedded vector zi
Descriptor is calculated, each overtrick is separately embedded；Select the spot net of a deep learning, in a network, input point
It is aligned by spatial alternation network, then by multilayer perceptron independent process, is finally summarized to indicate input shape first,
Middle input shape is a simple geometric object, indicates input shape using a small amount of point, and pass through a compact point net
Network executes insertion, is n by overtrick fast samplingp=128, to maintain effectively to calculate in batches and data to be promoted to increase, replace
Less than npOvertrick sampled, in principle from the point of view of this have no effect on the assessment in spot net maximum pond, however pass through and test table
It is bright：Less than or equal to nminp=40 overtrick can damage the overall performance of network in training, therefore set the insertion of overtrick to
Zero, make its semantic information that places one's entire reliance upon of classifying, in order to make spot net learn spatial distribution of different shapes, each overtrick is embedding
It is scaled to unit sphere before entering, using their normalization position pi', observation oiWith geometric properties fiTo indicate a little, to be
Coordination shape size, the raw metric diameter of overtrick is as the supplementary features after spot net maximum pond.
Wherein, the semantic segmentation is accumulated since the figure of overtrick is smaller than the figure established on original point cloud based on picture scroll
Deep learning algorithm can be classified to its node using abundant edge feature, to promote prolonged interaction to make
With.
Further, the segmentation, overtrick are embedded according to the information refinement that extreme edges is transmitted, specifically, each overtrick
SiThe hidden state being maintained in gating cycle unit (GRU) passes through embedded ziHidden state is initialized, is then used
Iteration t=1 ... T processing, in each iteration t, GRU is by its hidden stateWith an input informationAs defeated
Enter, and calculates its new hidden stateThe input information of overtrick iBy being used as hiding for adjacent overtrick j after calculating
StateWeighted sum, the practical weighting of extreme edges (j, i) depends on its attribute Fji, pass through the attribute of multilayer perceptron Θ
It calculates：
Wherein, ⊙ is Element-Level multiplication, and σ () is sigmoid function, and W. and b. are shared between all GRU train
Parameter, in equation (4)The update of expression standard GRU rules gates, in equation (5)It indicates to reset gate, in order to improve
Stability in training process, first by the input after linear transformation in equation (8)It is defined as ρ (a)：=(a-mean
(a))/(std (a)+∈) then converts the hidden state in equation (7)Wherein ∈ is a smaller constant.
Further, the gate is being inputted by equation (9)Gating information is hidden beforeGRU can be with
Input vector is reduced according to its hidden state, Θ returns a weight matrix to execute matrix-vector multiplication for each edge,
Although conventional convolution can be carried out by demonstrating it on grid, longer run time can be caused, occupy higher memory
And more parameters are generated, therefore a specific weight vector in edge is returned using equation (10), execute Element-Level multiplication.
Further, the hidden state, cascades hidden state, is connected in all time steps and hides shape
State, and linear transformation they generate parted pattern yi, given by following equation：
This makes in final classification, due to the increase of acceptance region, can utilize the dynamic of hidden state.
Wherein, the training and test, for training：It is embedded although the step of geometry is divided is unsupervised
Overtrick and semantic segmentation are by the way of intersecting entropy loss supervision, it is assumed that semantic nature is identical, and overtrick is being semantically uniform
, between the point that they are included, specify calibration label corresponding with most of labels, the Neighborhood Graph that match exponents is 3 carries out
Sampling, so that each SPG at most selections 512 are more than nminpOvertrick；For test：Complete mark SPG, in order to compensate for due to
Randomness caused by the sub-sampling of point cloud in spot net, the parted pattern obtained to 10 operations using different o'clock are put down
.
Fig. 2 is a kind of visible process figure of the large-scale point cloud semantic segmentation method based on overtrick figure of the present invention.Overtrick
The node of figure indicates simple shape, the edge then syntople between Expressive Features.Input point cloud (a) is divided into simple several
What shape, referred to as super point；(b) it is geometry division figure；On the basis of pretreated, connected in the extreme edges of each attribute attached
Close overtrick constructs overtrick figure (c)；Finally, it is compactly embedded in overtrick, using figure process of convolution peripheral information, and is classified
For semantic label, figure (d) indicates semantic segmentation.
Fig. 3 is a kind of segmentation instance graph of the large-scale point cloud semantic segmentation method based on overtrick figure of the present invention.By sweeping
It retouches desk and chair obtains the figure, figure (a) expression executes geometric zoning on cloud, then constructs overtrick figure (b), Mei Gechao
Grade o'clock executes insertion by a spot net, and network structure refines insertion in GRU shown in figure (c), and generates final label.
For those skilled in the art, the present invention is not limited to the details of above-described embodiment, in the essence without departing substantially from the present invention
In the case of refreshing and range, the present invention can be realized in other specific forms.In addition, those skilled in the art can be to this hair
Bright to carry out various modification and variations without departing from the spirit and scope of the present invention, these improvements and modifications also should be regarded as the present invention's
Protection domain.Therefore, the following claims are intended to be interpreted as including preferred embodiment and falls into all changes of the scope of the invention
More and change.
Claims (10)
1. a kind of large-scale point cloud semantic segmentation method based on overtrick figure, which is characterized in that include mainly geometry uniform segmentation
(1)；Overtrick figure builds (two)；Embedded overtrick (three)；Semantic segmentation (four)；Training and test (five).
2. based on the geometry uniform segmentation (one) described in claims 1, which is characterized in that a cloud is divided into geometry,
Referred to as overtrick will entirely put cloud as input using this unsupervised step, overtrick figure (SPG), SPG are calculated in geometric zoning
Each node correspond to the sub-fraction point cloud of geometrically simple object, it is contemplated that being semantically uniform, parameter passes through small
Point cloud is down-sampled to hundreds of points to indicate.
3. based on the subregion described in claims 2, which is characterized in that Universal Energy model can be divided and adapt to local geometric
Complexity, therefore input point cloud C is regarded as the point of one group of n three-dimensional point composition, by every using Universal Energy model come computational efficiency
A point i ∈ C define its position 3D pi, and other observed values are defined as o color or intensity etc.i, for each point, computational geometry
Feature dg,For characterizing the shape of its local neighborhood, variable sampling density is compensated by adaptive neighborhood, is utilized
Three linearity, flatness and scattering dimension values, and verticality is introduced, the elevation each put is calculated, after indicating normalization
piZ-axis coordinate on entirely input cloud, geometry uniform segmentation are the constant connection components of optimization problem reconciliation, are given by following formula
It is fixed：
Wherein [≠ 0] isFunction, whenWhen, [≠ 0] is equal to 0, in the case of other
Equal to 1, coefficient μ is regularization intensity, and for determining the roughness of institute's scoring area, constant communication component S={ S1,…,SkBe
The solution of equation (1), for defining geometrically simple element.
4. building (two) based on the overtrick figure described in claims 1, which is characterized in that SPG is the structured representation of a cloud,
Define a directional properties figureIts node is the set of overtrick S, the syntople extreme edges ε between overtrick
It indicates, with one group of dfFeature annotates extreme edges：Including the syntople between overtrick, by Gvor=(C, Evor) definition
For the symmetrical Voronoi adjacent maps of complete input point cloud, if EvorIn there are one edge, then S and T is two adjacent overtricks,
And S and T are located at both ends：
EvorTwo overtricks are connected, the significant correlation of extreme edges (S, T) is obtained from the edge offset amount δ (S, T) of the two overtricks
Feature：
δ (S, T)={ (pi-pj)|(i,j)∈Evor∩(S×T)} (3)
Extreme edges feature is exported by comparing the shapes and sizes of adjacent overtrick, is used | S | indicate the number for the point for including in overtrick S
Mesh, λ1,λ2,λ3Indicate each overtrick include point position covariance, by covariance export shape characteristic length (S)=
λ1, surface (S)=λ1λ2And volume (S)=λ1λ2λ3, by successively decreasing, sequence sorts.
5. based on the insertion overtrick (three) described in claims 1, which is characterized in that in each overtrick SiIn, select fixed size
Dimension dz, pass through embedded vector ziDescriptor is calculated, each overtrick is separately embedded；One deep learning of selection
Spot net, in a network, input point are aligned by spatial alternation network first, then by multilayer perceptron independent process, most
After summarize to indicate input shape, wherein input shape is a simple geometric object, a small amount of point is utilized to indicate input shape
Shape, and insertion is executed by a compact spot net, it is n by overtrick fast samplingp=128, to maintain effective batch
It calculates and data is promoted to increase, replace and be less than npOvertrick sampled, in principle from the point of view of this have no effect on spot net maximum pond
The assessment of change, however be shown experimentally that：Less than or equal to nminp=40 overtrick can damage the whole table of network in training
It is existing, therefore the insertion of overtrick is set as zero, make its semantic information that places one's entire reliance upon of classifying, in order to make spot net learn different shape
Spatial distribution, each overtrick is scaled to unit sphere before embedding, using their normalization position p 'i, observation oi
With geometric properties fiIndicate a little, in order to coordinate shape size, the raw metric diameter of overtrick as spot net maximum pondization it
Supplementary features afterwards.
6. based on the semantic segmentation (four) described in claims 1, which is characterized in that since the figure of overtrick is than on original point cloud
The figure of foundation is small, can be divided its node using abundant edge feature based on the deep learning algorithm of picture scroll product
Class, to promote prolonged reciprocation.
7. based on the segmentation described in claims 6, which is characterized in that overtrick is embedded according to the information refinement that extreme edges is transmitted,
Specifically, each overtrick SiThe hidden state being maintained in gating cycle unit (GRU) passes through embedded ziHidden state is carried out
Then initialization uses iteration t=1 ... T processing, in each iteration t, GRU is by its hidden stateIt is inputted with one
InformationAs input, and calculate its new hidden stateThe input information of overtrick iBy being used as phase after calculating
The hidden state of adjacent overtrick jWeighted sum, the practical weighting of extreme edges (j, i) depends on its attribute Fji, pass through multilayer sense
Know that the attribute of device Θ calculates：
Wherein, ⊙ is Element-Level multiplication, and σ () is sigmoid function, W. and b. be shared between all GRU can training parameter,
In equation (4)The update of expression standard GRU rules gates, r in equation (5)i (t)It indicates to reset gate, in order to improve training
Stability in the process, first by the input after linear transformation in equation (8)It is defined as ρ (a)：=(a-mean (a))/
(std (a)+∈) then converts the hidden state in equation (7)Wherein ∈ is a smaller constant.
8. based on the gate described in claims 7, which is characterized in that inputted by equation (9)Gate letter is hidden before
BreathGRU can reduce input vector according to its hidden state, and Θ returns a weight matrix to be executed for each edge
Matrix-vector multiplication can lead to longer run time, account for although conventional convolution can be carried out by demonstrating it on grid
With higher memory and more parameters are generated, therefore a specific weight vector in edge is returned using equation (10), are held
Row element grade multiplication.
9. based on the hidden state described in claims 8, which is characterized in that cascaded to hidden state, in institute's having time
Connect hidden state in step, and linear transformation they generate parted pattern yi, given by following equation：
This makes in final classification, due to the increase of acceptance region, can utilize the dynamic of hidden state.
10. based on described in claims 1 training and test (five), which is characterized in that for training：Although geometry segmentation
Step is unsupervised, but embedded overtrick and semantic segmentation are by the way of intersecting entropy loss supervision, it is assumed that semantic nature
Identical, overtrick is being semantically uniform, between the point that they are included, specifies calibration label opposite with most of labels
It answers, the Neighborhood Graph that match exponents is 3 is sampled, so that each SPG at most selections 512 are more than nminpOvertrick；For test：
Complete mark SPG, in order to compensate for randomness caused by the sub-sampling due to the point cloud in spot net, using different o'clock to 10 times
The parted pattern that operation obtains is averaged.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810132821.9A CN108319957A (en) | 2018-02-09 | 2018-02-09 | A kind of large-scale point cloud semantic segmentation method based on overtrick figure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810132821.9A CN108319957A (en) | 2018-02-09 | 2018-02-09 | A kind of large-scale point cloud semantic segmentation method based on overtrick figure |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108319957A true CN108319957A (en) | 2018-07-24 |
Family
ID=62903904
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810132821.9A Pending CN108319957A (en) | 2018-02-09 | 2018-02-09 | A kind of large-scale point cloud semantic segmentation method based on overtrick figure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108319957A (en) |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109598727A (en) * | 2018-11-28 | 2019-04-09 | 北京工业大学 | A kind of CT image pulmonary parenchyma three-dimensional semantic segmentation method based on deep neural network |
| CN109858374A (en) * | 2018-12-31 | 2019-06-07 | 武汉中海庭数据技术有限公司 | Arrow class graticule extraction method and device in high-precision cartography |
| CN110059608A (en) * | 2019-04-11 | 2019-07-26 | 腾讯科技（深圳）有限公司 | A kind of object detecting method, device, electronic equipment and storage medium |
| CN110097556A (en) * | 2019-04-29 | 2019-08-06 | 东南大学 | Large-scale point cloud semantic segmentation algorithm based on PointNet |
| CN110197215A (en) * | 2019-05-22 | 2019-09-03 | 深圳市牧月科技有限公司 | A kind of ground perception point cloud semantic segmentation method of autonomous driving |
| CN110210431A (en) * | 2019-06-06 | 2019-09-06 | 上海黑塞智能科技有限公司 | A kind of point cloud classifications method based on cloud semantic tagger and optimization |
| CN110569851A (en) * | 2019-08-28 | 2019-12-13 | 广西师范大学 | real-time semantic segmentation method for gated multi-layer fusion |
| CN110827302A (en) * | 2019-11-14 | 2020-02-21 | 中南大学 | Point cloud target extraction method and device based on depth map convolutional network |
| CN110838122A (en) * | 2018-08-16 | 2020-02-25 | 北京大学 | Point cloud segmentation method and device and computer storage medium |
| CN110958455A (en) * | 2018-09-26 | 2020-04-03 | 华为技术有限公司 | Point cloud coding and decoding method and coder-decoder |
| CN111027559A (en) * | 2019-10-31 | 2020-04-17 | 湖南大学 | Point cloud semantic segmentation method based on expansion point convolution space pyramid pooling |
| CN111583263A (en) * | 2020-04-30 | 2020-08-25 | 北京工业大学 | Point cloud segmentation method based on joint dynamic graph convolution |
| CN111860768A (en) * | 2020-06-16 | 2020-10-30 | 中山大学 | Method for enhancing point-edge interaction of graph neural network |
| CN111968121A (en) * | 2020-08-03 | 2020-11-20 | 电子科技大学 | Three-dimensional point cloud scene segmentation method based on instance embedding and semantic fusion |
| CN112241676A (en) * | 2020-07-07 | 2021-01-19 | 西北农林科技大学 | Method for automatically identifying terrain sundries |
| CN112598805A (en) * | 2020-12-24 | 2021-04-02 | 浙江商汤科技开发有限公司 | Prompt message display method, device, equipment and storage medium |
| CN112785611A (en) * | 2021-01-29 | 2021-05-11 | 昆明理工大学 | 3D point cloud weak supervision semantic segmentation method and system |
| CN113484875A (en) * | 2021-07-30 | 2021-10-08 | 燕山大学 | Laser radar point cloud target hierarchical identification method based on mixed Gaussian ordering |
| CN114401666A (en) * | 2019-07-15 | 2022-04-26 | 普罗马顿控股有限责任公司 | Object detection and instance segmentation of 3D point clouds based on deep learning |
| CN114743010A (en) * | 2022-06-13 | 2022-07-12 | 山东科技大学 | Ultrahigh voltage power transmission line point cloud data semantic segmentation method based on deep learning |
| CN115641583A (en) * | 2022-12-26 | 2023-01-24 | 苏州赫芯科技有限公司 | Point cloud detection method, system and medium based on self-supervision and active learning |
| CN115731137A (en) * | 2022-11-29 | 2023-03-03 | 湖北大学 | Outdoor large scene point cloud segmentation method based on A-EdgeConv |
| CN117392386A (en) * | 2023-10-13 | 2024-01-12 | 浙江省测绘科学技术研究院 | Classification training method and device for superside mask generation network based on instance segmentation |
| US11967873B2 (en) | 2019-09-23 | 2024-04-23 | Canoo Technologies Inc. | Fractional slot electric motors with coil elements having rectangular cross-sections |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103914875A (en) * | 2014-04-17 | 2014-07-09 | 中国科学院深圳先进技术研究院 | Functional modeling method for indoor scene |
| CN105488809A (en) * | 2016-01-14 | 2016-04-13 | 电子科技大学 | Indoor scene meaning segmentation method based on RGBD descriptor |
-
2018
- 2018-02-09 CN CN201810132821.9A patent/CN108319957A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103914875A (en) * | 2014-04-17 | 2014-07-09 | 中国科学院深圳先进技术研究院 | Functional modeling method for indoor scene |
| CN105488809A (en) * | 2016-01-14 | 2016-04-13 | 电子科技大学 | Indoor scene meaning segmentation method based on RGBD descriptor |
Non-Patent Citations (1)
| Title |
|---|
| LOIC LANDRIEU ET.AL: "Large-scale Point Cloud Semantic Segmentation with Superpoint Graphs", 《ARXIV:1711.09869V1 [CS.CV]》 * |
Cited By (34)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110838122B (en) * | 2018-08-16 | 2022-06-14 | 北京大学 | Point cloud segmentation method and device and computer storage medium |
| CN110838122A (en) * | 2018-08-16 | 2020-02-25 | 北京大学 | Point cloud segmentation method and device and computer storage medium |
| CN110958455A (en) * | 2018-09-26 | 2020-04-03 | 华为技术有限公司 | Point cloud coding and decoding method and coder-decoder |
| CN109598727A (en) * | 2018-11-28 | 2019-04-09 | 北京工业大学 | A kind of CT image pulmonary parenchyma three-dimensional semantic segmentation method based on deep neural network |
| CN109598727B (en) * | 2018-11-28 | 2021-09-14 | 北京工业大学 | CT image lung parenchyma three-dimensional semantic segmentation method based on deep neural network |
| CN109858374A (en) * | 2018-12-31 | 2019-06-07 | 武汉中海庭数据技术有限公司 | Arrow class graticule extraction method and device in high-precision cartography |
| CN109858374B (en) * | 2018-12-31 | 2021-08-17 | 武汉中海庭数据技术有限公司 | Automatic extraction method and device for arrow mark lines in high-precision map making |
| CN110059608B (en) * | 2019-04-11 | 2021-07-06 | 腾讯科技（深圳）有限公司 | Object detection method and device, electronic equipment and storage medium |
| CN110059608A (en) * | 2019-04-11 | 2019-07-26 | 腾讯科技（深圳）有限公司 | A kind of object detecting method, device, electronic equipment and storage medium |
| CN110097556A (en) * | 2019-04-29 | 2019-08-06 | 东南大学 | Large-scale point cloud semantic segmentation algorithm based on PointNet |
| CN110197215A (en) * | 2019-05-22 | 2019-09-03 | 深圳市牧月科技有限公司 | A kind of ground perception point cloud semantic segmentation method of autonomous driving |
| CN110210431A (en) * | 2019-06-06 | 2019-09-06 | 上海黑塞智能科技有限公司 | A kind of point cloud classifications method based on cloud semantic tagger and optimization |
| CN114401666A (en) * | 2019-07-15 | 2022-04-26 | 普罗马顿控股有限责任公司 | Object detection and instance segmentation of 3D point clouds based on deep learning |
| CN110569851A (en) * | 2019-08-28 | 2019-12-13 | 广西师范大学 | real-time semantic segmentation method for gated multi-layer fusion |
| CN110569851B (en) * | 2019-08-28 | 2022-03-15 | 广西师范大学 | Real-time semantic segmentation method for gated multi-layer fusion |
| US11967873B2 (en) | 2019-09-23 | 2024-04-23 | Canoo Technologies Inc. | Fractional slot electric motors with coil elements having rectangular cross-sections |
| CN111027559A (en) * | 2019-10-31 | 2020-04-17 | 湖南大学 | Point cloud semantic segmentation method based on expansion point convolution space pyramid pooling |
| CN110827302A (en) * | 2019-11-14 | 2020-02-21 | 中南大学 | Point cloud target extraction method and device based on depth map convolutional network |
| CN111583263A (en) * | 2020-04-30 | 2020-08-25 | 北京工业大学 | Point cloud segmentation method based on joint dynamic graph convolution |
| CN111583263B (en) * | 2020-04-30 | 2022-09-23 | 北京工业大学 | Point cloud segmentation method based on joint dynamic graph convolution |
| CN111860768A (en) * | 2020-06-16 | 2020-10-30 | 中山大学 | Method for enhancing point-edge interaction of graph neural network |
| CN111860768B (en) * | 2020-06-16 | 2023-06-09 | 中山大学 | Method for enhancing point-edge interaction of graph neural network |
| CN112241676A (en) * | 2020-07-07 | 2021-01-19 | 西北农林科技大学 | Method for automatically identifying terrain sundries |
| CN111968121B (en) * | 2020-08-03 | 2021-12-03 | 电子科技大学 | Three-dimensional point cloud scene segmentation method based on instance embedding and semantic fusion |
| CN111968121A (en) * | 2020-08-03 | 2020-11-20 | 电子科技大学 | Three-dimensional point cloud scene segmentation method based on instance embedding and semantic fusion |
| CN112598805A (en) * | 2020-12-24 | 2021-04-02 | 浙江商汤科技开发有限公司 | Prompt message display method, device, equipment and storage medium |
| CN112785611A (en) * | 2021-01-29 | 2021-05-11 | 昆明理工大学 | 3D point cloud weak supervision semantic segmentation method and system |
| CN112785611B (en) * | 2021-01-29 | 2021-10-12 | 昆明理工大学 | 3D point cloud weak supervision semantic segmentation method and system |
| CN113484875A (en) * | 2021-07-30 | 2021-10-08 | 燕山大学 | Laser radar point cloud target hierarchical identification method based on mixed Gaussian ordering |
| CN114743010A (en) * | 2022-06-13 | 2022-07-12 | 山东科技大学 | Ultrahigh voltage power transmission line point cloud data semantic segmentation method based on deep learning |
| CN114743010B (en) * | 2022-06-13 | 2022-08-26 | 山东科技大学 | Ultrahigh voltage power transmission line point cloud data semantic segmentation method based on deep learning |
| CN115731137A (en) * | 2022-11-29 | 2023-03-03 | 湖北大学 | Outdoor large scene point cloud segmentation method based on A-EdgeConv |
| CN115641583A (en) * | 2022-12-26 | 2023-01-24 | 苏州赫芯科技有限公司 | Point cloud detection method, system and medium based on self-supervision and active learning |
| CN117392386A (en) * | 2023-10-13 | 2024-01-12 | 浙江省测绘科学技术研究院 | Classification training method and device for superside mask generation network based on instance segmentation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108319957A (en) | A kind of large-scale point cloud semantic segmentation method based on overtrick figure | |
| Bao et al. | A novel hybrid harris hawks optimization for color image multilevel thresholding segmentation | |
| CN109993748B (en) | Three-dimensional grid object segmentation method based on point cloud processing network | |
| Li et al. | Building-a-nets: Robust building extraction from high-resolution remote sensing images with adversarial networks | |
| CN105243139B (en) | A kind of method for searching three-dimension model and its retrieval device based on deep learning | |
| CN105809672B (en) | A kind of image multiple target collaboration dividing method constrained based on super-pixel and structuring | |
| CN107066559A (en) | A kind of method for searching three-dimension model based on deep learning | |
| Lokanath et al. | Accurate object classification and detection by faster-RCNN | |
| CN106650744B (en) | The image object of local shape migration guidance is divided into segmentation method | |
| CN111625667A (en) | Three-dimensional model cross-domain retrieval method and system based on complex background image | |
| Wang et al. | GKFC-CNN: Modified Gaussian kernel fuzzy C-means and convolutional neural network for apple segmentation and recognition | |
| CN111915629B (en) | Super-pixel segmentation method based on boundary detection | |
| CN105354593B (en) | A kind of threedimensional model sorting technique based on NMF | |
| CN109711401A (en) | A kind of Method for text detection in natural scene image based on Faster Rcnn | |
| CN108846404A (en) | A kind of image significance detection method and device based on the sequence of related constraint figure | |
| CN104063721B (en) | A kind of human behavior recognition methods learnt automatically based on semantic feature with screening | |
| CN112784782B (en) | Three-dimensional object identification method based on multi-view double-attention network | |
| Yan et al. | Monocular depth estimation with guidance of surface normal map | |
| CN106844739A (en) | A kind of Remote Sensing Imagery Change information retrieval method based on neutral net coorinated training | |
| Chen et al. | Agricultural remote sensing image cultivated land extraction technology based on deep learning | |
| Huang et al. | Hybrid bio-inspired lateral inhibition and imperialist competitive algorithm for complicated image matching | |
| CN108009557A (en) | A kind of threedimensional model method for describing local characteristic based on shared weight convolutional network | |
| Mishra et al. | Medical image retrieval using self-organising map on texture features | |
| CN110197206A (en) | The method and device of image procossing | |
| CN111680579A (en) | Remote sensing image classification method for adaptive weight multi-view metric 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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180724 |
|
| WD01 | Invention patent application deemed withdrawn after publication |