CN110176078A - A kind of mask method and device of training set data - Google Patents
A kind of mask method and device of training set data Download PDFInfo
- Publication number
- CN110176078A CN110176078A CN201910443047.8A CN201910443047A CN110176078A CN 110176078 A CN110176078 A CN 110176078A CN 201910443047 A CN201910443047 A CN 201910443047A CN 110176078 A CN110176078 A CN 110176078A
- Authority
- CN
- China
- Prior art keywords
- point cloud
- cloud data
- laser point
- marked
- callout box
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T19/00—Manipulating 3D models or images for computer graphics
- G06T19/20—Editing of 3D images, e.g. changing shapes or colours, aligning objects or positioning parts
Landscapes
- Engineering & Computer Science (AREA)
- Architecture (AREA)
- Computer Graphics (AREA)
- Computer Hardware Design (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Length Measuring Devices By Optical Means (AREA)
Abstract
The embodiment of the present invention discloses the mask method and device of a kind of training set data.This method comprises: being trained to target network model according to the sample laser point cloud data and standard callout box that have marked, obtaining updated network model；When the quantity for the sample laser point cloud data not marked in data set to be marked is greater than the first quantity, never the second sample laser point cloud data is determined in the sample laser point cloud data marked, and the reference callout box of object to be marked in the second sample laser point cloud data is determined by updated network model；It is shown in two dimension vertical view interface according to the first mapping relations by the second sample laser point cloud data and with reference to callout box, is operated according to mark person for the first adjustment with reference to callout box, determine standard callout box；The training set is added in second sample laser point cloud data and corresponding standard callout box；Repeat the above process.Using scheme provided in an embodiment of the present invention, the annotating efficiency of training set data can be improved.
Description
Technical field
The present invention relates to intelligent driving technical fields, in particular to the mask method and dress of a kind of training set data
It sets.
Background technique
In machine learning field, training set data contains the great amount of samples data for training network model.To instruction
Practicing collection data to be labeled is a ring indispensable in current artificial intelligence industry, and the work of data mark is more both at home and abroad at present
Using the artificial labelling schemes such as crowdsourcing, outsourcing.Above scheme needs that a large amount of human resources is employed just to be able to satisfy in machine learning
Mark demand in continuous iterative process, inefficiency.
When data to be marked are laser point cloud data, the process manually marked will be more complicated.Laser point cloud data
It is the data type that laser radar returns, laser point cloud includes a large amount of laser point datas.Laser radar may be mounted at intelligent vehicle
, in the equipment such as robot, for that can be determined according to laser point cloud data according to the laser point cloud data around acquisition equipment
Obstacle information around equipment.
Collected laser point cloud data information content usually it is excessive excessively miscellaneous, it is difficult to directly use, need manually to its into
Row identification and mark.In practical operation, it usually needs mark out the three-dimensional of object in Three Dimensional Interface and surround frame.It is vertical in mark
It is relatively difficult when body encirclement frame, it needs to mark personnel and repeatedly adjusts.In face of a large amount of, numerous and jumbled laser point cloud data, this logarithm
It is lower according to the mode efficiency of mark.In order to solve the efficiency in this training set data mark, a kind of raising training is needed
Collect the mask method of the annotating efficiency of data.
Summary of the invention
The present invention provides a kind of mask method of training set data and devices, to improve the mark effect of training set data
Rate.Specific technical solution is as follows.
In a first aspect, the embodiment of the invention discloses a kind of mask methods of training set data, comprising:
Obtain the standard mark of the first sample laser point cloud data and corresponding object to be marked that have marked in training set
Frame, as model training data；
According to the model training data, target network model is trained, obtains updated network model；Its
In, the updated network model is used for so that sample laser point cloud data and corresponding standard callout box are interrelated；
Judge whether the quantity for the sample laser point cloud data not marked in data set to be marked is greater than the first quantity；
If it is greater, then determining the second quantity from the sample laser point cloud data not marked in the data set to be marked
A sample laser point cloud data determines second sample by updated network model as the second sample laser point cloud data
The reference callout box of object to be marked in this laser point cloud data；
The first mapping relations between coordinate system and three-dimensional system of coordinate are overlooked according to two dimension, by the second sample laser point
Cloud data and reference callout box are shown in two dimension and overlook in interface；Wherein, the three-dimensional system of coordinate is the second sample laser
Coordinate system where point cloud data, it is corresponding with the two dimension vertical view coordinate system that the two dimension overlooks interface；
Mark person is obtained to overlook the second sample laser point cloud data of interface display for the two dimension and refer to callout box
The first adjustment of input operates, and is operated, is determined to be marked in the second sample laser point cloud data according to the first adjustment
The standard callout box of object；
The training set is added in the second sample laser point cloud data and corresponding standard callout box；By described second
Sample laser point cloud data and corresponding standard callout box are as model training data, using updated network model as target
Network model, return execution is described to be trained target network model according to the model training data, obtains updated
The step of network model.
Optionally, in determining the second sample laser point cloud data after the reference callout box of object to be marked, institute
State method further include:
The second sample laser point cloud data is shown in Three Dimensional Interface；Wherein, the Three Dimensional Interface and described three
It is corresponding to tie up coordinate system；
It is described to be operated according to the first adjustment, determine the mark of object to be marked in the second sample laser point cloud data
The step of quasi- callout box, comprising:
Operated according to the first adjustment, determine first of object to be marked in the second sample laser point cloud data to
Adjust callout box；
The described first callout box to be adjusted is shown in the Three Dimensional Interface；
The second adjustment that mark person is obtained for the first callout box input to be adjusted shown in the Three Dimensional Interface operates,
It is operated according to the second adjustment, the described first callout box to be adjusted is adjusted, the second sample laser point cloud is obtained
The standard callout box of object to be marked in data.
Optionally, described to be operated according to the first adjustment, it determines to be marked in the second sample laser point cloud data
The step of standard callout box of object, comprising:
Operated according to the first adjustment, determine second of object to be marked in the second sample laser point cloud data to
Adjust callout box；
Determine other two-dimensional coordinate systems and the three-dimensional system of coordinate where the other faces of the described second callout box to be adjusted
Between the second mapping relations；Wherein, the other faces include the back side and/or side, other described two-dimensional coordinate systems include: two
Dimension back apparent coordinates system and/or two-dimentional side view coordinate system；
According to second mapping relations, by the second sample laser point cloud data and second callout box to be adjusted
It is shown in other two-dimentional interfaces corresponding with other described two-dimensional coordinate systems；
The third for obtaining second to be adjusted callout box input of the mark person for other two-dimentional interface displays adjusts behaviour
Make；
The described second callout box to be adjusted is adjusted according to third adjustment operation, second sample is obtained and swashs
The standard callout box of object to be marked in light point cloud data.
Optionally, when updated network model determines that there is no to be marked right in the second sample laser point cloud data
As when, the method also includes:
The second sample laser point cloud data is added in negative sample training set；Sample in the negative sample training set
Laser point cloud data refusal is shown for mark person.
Optionally, described according to the model training data, the step of being trained to target network model, comprising:
It will be in the sample laser point cloud data input target network model in the model training data；The target network
Model includes feature extraction layer and recurrence layer；
By the first model parameter of the feature extraction layer, determine feature in the sample laser point cloud data to
Amount；By second model parameter for returning layer, described eigenvector is returned, initial callout box is obtained；
Determine initial callout box standard mark corresponding with sample laser point cloud data in the model training data
Difference between frame；
When the difference is greater than default discrepancy threshold, first model parameter and described the are modified according to the difference
Two model parameters return and execute the step inputted the sample laser point cloud data in target network model；
When the difference is not more than default discrepancy threshold, determine that the target network model training is completed.
Optionally, when the quantity for the sample laser point cloud data not marked in the data set to be marked is no more than the first number
When amount, the method also includes:
Using the sample laser point cloud data not marked in the data set to be marked as third sample laser point cloud data,
Directly display the third sample laser point cloud data；
Obtain the labeling operation that mark person is directed to third sample laser point cloud data input；
According to the labeling operation, the standard mark of the object to be marked for the third sample laser point cloud data is determined
Infuse frame；
The training set is added in the third sample laser point cloud data and corresponding standard callout box.
Second aspect, the embodiment of the invention discloses a kind of annotation equipments of training set data, comprising:
Data acquisition module is configured as obtaining in training set the first sample laser point cloud data that has marked and corresponding
The standard callout box of object to be marked, as model training data；
Model training module is configured as being trained target network model according to the model training data, obtaining
Updated network model；Wherein, the updated network model is used for so that sample laser point cloud data and corresponding mark
Quasi- callout box is interrelated；
Quantity judgment module is configured as the quantity for the sample laser point cloud data for judging not mark in data set to be marked
Whether the first quantity is greater than；
Reference block determining module is configured as the sample laser point cloud data not marked in the data set to be marked
When quantity is greater than the first quantity, the second quantity is determined from the sample laser point cloud data not marked in the data set to be marked
A sample laser point cloud data determines second sample by updated network model as the second sample laser point cloud data
The reference callout box of object to be marked in this laser point cloud data；
Two-dimentional display module, the first mapping for being configured as being overlooked between coordinate system and three-dimensional system of coordinate according to two dimension are closed
System shows in two dimension vertical view interface by the second sample laser point cloud data and with reference to callout box；Wherein, the three-dimensional seat
Mark system is the coordinate system where the second sample laser point cloud data, and the two dimension overlooks interface and the two dimension overlooks coordinate
System corresponds to；
Template determining module is configured as the second sample that acquisition mark person overlooks interface display for the two dimension and swashs
Light point cloud data and the first adjustment operation inputted with reference to callout box determine second sample according to the first adjustment operation
The standard callout box of object to be marked in this laser point cloud data；
First is added module, is configured as the second sample laser point cloud data and corresponding standard callout box being added
The training set；
Data update module, be configured as using the second sample laser point cloud data and corresponding standard callout box as
It is described according to the model training to return to execution using updated network model as target network model for model training data
Data are trained target network model, obtain the operation of updated network model.
Optionally, described device further include:
Three-dimensional Display module is configured as the reference of the object to be marked in determining the second sample laser point cloud data
After callout box, the second sample laser point cloud data is shown in Three Dimensional Interface；Wherein, the Three Dimensional Interface with it is described
Three-dimensional system of coordinate is corresponding；
The template determining module operates according to the first adjustment, determines the second sample laser point cloud data
In object to be marked standard callout box when, comprising:
Operated according to the first adjustment, determine first of object to be marked in the second sample laser point cloud data to
Adjust callout box；
The described first callout box to be adjusted is shown in the Three Dimensional Interface；
The second adjustment that mark person is obtained for the first callout box input to be adjusted shown in the Three Dimensional Interface operates,
It is operated according to the second adjustment, the described first callout box to be adjusted is adjusted, the second sample laser point cloud is obtained
The standard callout box of object to be marked in data.
Optionally, the template determining module operates according to the first adjustment, determines the second sample laser point
In cloud data when the standard callout box of object to be marked, comprising:
Operated according to the first adjustment, determine second of object to be marked in the second sample laser point cloud data to
Adjust callout box；
Determine other two-dimensional coordinate systems and the three-dimensional system of coordinate where the other faces of the described second callout box to be adjusted
Between the second mapping relations；Wherein, the other faces include the back side and/or side, other described two-dimensional coordinate systems include: two
Dimension back apparent coordinates system and/or two-dimentional side view coordinate system；
According to second mapping relations, by the second sample laser point cloud data and second callout box to be adjusted
It is shown in other two-dimentional interfaces corresponding with other described two-dimensional coordinate systems；
The third for obtaining second to be adjusted callout box input of the mark person for other two-dimentional interface displays adjusts behaviour
Make；
The described second callout box to be adjusted is adjusted according to third adjustment operation, second sample is obtained and swashs
The standard callout box of object to be marked in light point cloud data.
Optionally, described device further include:
Second is added module, is configured as determining in the second sample laser point cloud data when updated network model
There is no when object to be marked, the second sample laser point cloud data is added in negative sample training set；The negative sample instruction
Practice the sample laser point cloud data refusal concentrated to be shown for mark person.
Optionally, the model training module, is specifically configured to:
It will be in the sample laser point cloud data input target network model in the model training data；The target network
Model includes feature extraction layer and recurrence layer；
By the first model parameter of the feature extraction layer, determine feature in the sample laser point cloud data to
Amount；By second model parameter for returning layer, described eigenvector is returned, initial callout box is obtained；
Determine initial callout box standard mark corresponding with sample laser point cloud data in the model training data
Difference between frame；
When the difference is greater than default discrepancy threshold, first model parameter and described the are modified according to the difference
Two model parameters return and execute the operation inputted the sample laser point cloud data in target network model；
When the difference is not more than default discrepancy threshold, determine that the target network model training is completed.
Optionally, the two-dimentional display module, is additionally configured to the sample not marked in the data set to be marked and swashs
When the quantity of light point cloud data is not more than the first quantity, the sample laser point cloud data that will not be marked in the data set to be marked
As third sample laser point cloud data, the third sample laser point cloud data is directly displayed；
It is defeated for the third sample laser point cloud data to be additionally configured to acquisition mark person for the template determining module
The labeling operation entered；According to the labeling operation, the object to be marked for being directed to the third sample laser point cloud data is determined
Standard callout box；
Described first is added module, is additionally configured to mark the third sample laser point cloud data and corresponding standard
The training set is added in frame.
As shown in the above, the mask method and device of training set data provided in an embodiment of the present invention, can will
The sample laser point cloud data of mark and corresponding standard callout box are trained target network model, use the net after training
Network model determines reference the callout box of object to be marked in the laser point cloud data that does not mark, using with reference to callout box as referring to,
It is labeled for mark person, obtains standard callout box, can reduce the mark complexity of mark person in this way, improve training set data
Annotating efficiency；Simultaneously as laser point cloud data is the data being distributed in three-dimensional space, overlooks in interface and show in two dimension
Laser point cloud data and reference callout box are adjusted reference callout box for mark person, and then obtain standard callout box, instead of
Originally directly mark is three-dimensional in three dimensions surrounds frame, overlooks in interface in two dimension and is adjusted to reference callout box, can
Difficulty when mark person's mark is largely reduced, the annotating efficiency of training set data is improved.Certainly, implement the present invention
Any product or method do not necessarily require achieving all the advantages described above at the same time.
The innovative point of the embodiment of the present invention includes:
1, network model is trained according to the data marked, using the model after training to the data not marked
Be labeled, and two dimension overlook interface in show laser point cloud data and model determination callout box, be supplied to mark person into
Row adjustment, obtains standard callout box, and the case where for great amount of samples data and sample data is this mark hardly possible of laser point cloud
It spends for biggish data, this notation methods can reduce mark difficulty, improve the efficiency of mark.
2, it when determining standard callout box, can also will be shown with reference to callout box in other two-dimentional interfaces, for mark person
The three-dimensional other faces for surrounding frame are adjusted, more accurate standard callout box can be obtained, compared to directly in Three Dimensional Interface
In other faces are adjusted, annotating efficiency can be turned up.
3, when network model determines, and object to be marked is not present in sample laser point cloud data, no longer by the sample laser
Point cloud data is shown to mark person and is identified, but is directly added into negative sample training set, uses for other model trainings.When
There are when great amount of samples data, such operation can screen sample data, and the laser of object to be marked only will be present
Point cloud data is supplied to mark person and is labeled, and can be improved the efficiency of mark.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described.It should be evident that the accompanying drawings in the following description is only this hair
Bright some embodiments.It for those of ordinary skill in the art, without creative efforts, can be with root
Other attached drawings are obtained according to these attached drawings.
Fig. 1 is a kind of flow diagram of the mask method of training set data provided in an embodiment of the present invention；
Fig. 2 is a kind of coordinate frame system schematic diagram provided in an embodiment of the present invention；
Fig. 3 is a kind of screenshot capture of display interface provided in an embodiment of the present invention；
Fig. 4 is a kind of signal that view, side view coordinate system were overlooked, carried on the back to three-dimensional system of coordinate provided in an embodiment of the present invention and two dimension
Figure；
Fig. 5 is a kind of execution flow diagram of embodiment provided by the invention；
Fig. 6 is the logical schematic provided in an embodiment of the present invention screened to sample laser point cloud data；
Fig. 7 is a kind of structural schematic diagram of the annotation equipment of training set data provided in an embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Whole description.Obviously, described embodiment is only a part of the embodiments of the present invention, instead of all the embodiments.Base
Embodiment in the present invention, those of ordinary skill in the art under that premise of not paying creative labor it is obtained it is all its
His embodiment, shall fall within the protection scope of the present invention.
It should be noted that term " includes " and " having " and their any change in the embodiment of the present invention and attached drawing
Shape, it is intended that cover and non-exclusive include.Such as comprising the process, method of a series of steps or units, system, product or
Equipment is not limited to listed step or unit, but optionally further comprising the step of not listing or unit or optional
Ground further includes other step or units intrinsic for these process, methods, product or equipment.
The embodiment of the invention discloses a kind of mask method of training set data and devices, can be improved training set data
Annotating efficiency.The embodiment of the present invention is described in detail below.
Fig. 1 is a kind of flow diagram of the mask method of training set data provided in an embodiment of the present invention.This method is answered
For electronic equipment.The electronic equipment can be common computer, server or Intelligent mobile equipment etc..This method specifically includes
Following steps.
S110: the standard of the first sample laser point cloud data and corresponding object to be marked that have marked in training set is obtained
Callout box, as model training data.
Wherein, training set includes for training multiple sample laser point cloud datas of network model and corresponding to be marked right
The standard callout box of elephant.Each sample laser point cloud data includes multiple laser data points.Sample laser point cloud data can be
What laser radar collected.Laser radar can be installed in smart machine, and smart machine can be intelligent vehicle, machine
The equipment such as people.Laser radar can acquire the laser point cloud data centered on laser radar, in one ring border of surrounding.
Laser radar emits multiple laser beams when acquiring data, to ambient enviroment, and each laser beam encounters object reflection
Return laser radar.Laser radar is according to each laser beam of transmitting and the available laser point cloud number of each laser beam of return
According to laser point cloud data can characterize the three-dimensional space position of the surrounding objects centered on laser radar.
Object to be marked may include vehicle, pedestrian, bicycle, tricycle etc..Standard callout box can be understood as can
It surrounds the three-dimensional of object to be marked and surrounds frame, standard callout box can be indicated with rectangular parallelepiped form.First sample laser point cloud number
According to quantity can may include multiple objects to be marked in each first laser point cloud data to be multiple, it is each to be marked right
As a corresponding standard callout box.
S120: according to above-mentioned model training data, being trained target network model, obtains updated network mould
Type.
Wherein, updated network model is used for so that sample laser point cloud data is mutually closed with corresponding standard callout box
Connection.
In this step, target network model can use deep learning network model.The obtained network model of training, can be with
According to the callout box of object to be detected in the laser point cloud data of the determining input of model parameter, which be can be used as with reference to mark
Note.Target network model is trained according to above-mentioned model training data, obtained network model has certain accuracy,
Sample laser point cloud data and corresponding standard callout box can be made interrelated to a certain extent.
S130: judge whether the quantity for the sample laser point cloud data not marked in data set to be marked is greater than the first number
Amount, if it is greater, then executing step S140.
Data set to be marked includes the sample laser point cloud data not marked largely.When what is do not marked in data set to be marked
When the quantity of sample laser point cloud data is not more than the first quantity, it may not need and first determined using network model with reference to callout box,
Mark person is supplied directly to be labeled sample laser point cloud data.
When the quantity for the sample laser point cloud data not marked in data set to be marked is greater than the first quantity, in order to improve
The efficiency of mark can execute step S140, improve annotating efficiency.
S140: the second quantity sample laser is determined from the sample laser point cloud data not marked in data set to be marked
Point cloud data determines the second sample laser point cloud data by updated network model as the second sample laser point cloud data
In object to be marked reference callout box.
Wherein, the second quantity can also be greater than the first quantity less than the first quantity.For example, the first quantity can be
2000, the second quantity can be 1000.When the quantity for the sample laser point cloud data not marked in data set to be marked is 5000
When, 1000 sample laser point cloud datas can be chosen from this 5000 sample laser point cloud datas as the second sample laser
Point cloud data.
Each second sample laser point cloud data is inputted into updated network model, by updated network model according to
The obtained network parameter of network model training determines that reference of object to be marked in each second sample laser point cloud data is marked
Infuse frame.Wherein, it is three-dimensional encirclement frame with reference to callout box, can be indicated using cuboid.This may be not enough with reference to callout box
Accurately, the person of mark is needed to be adjusted reference callout box, to improve the accuracy of callout box.
S150: the first mapping relations between coordinate system and three-dimensional system of coordinate are overlooked according to two dimension, by the second sample laser
Point cloud data and reference callout box are shown in two dimension and overlook in interface.
Wherein, three-dimensional system of coordinate is the coordinate system where the second sample laser point cloud data, and three-dimensional system of coordinate can be to swash
Point in optical radar is coordinate origin, using the direction of advance of smart machine as X-direction, with straight up for Z-direction, and intelligence
The front-left of equipment is to for Y direction.
It is corresponding with two dimension vertical view coordinate system that two dimension overlooks interface.Since the direction of Z axis in above-mentioned three-dimensional system of coordinate is vertical
Upwardly direction, therefore it can be the two-dimensional coordinate system comprising X-axis in three-dimensional cartesian coordinate system and Y-axis that two dimension, which overlooks coordinate system,.X
Axis and Y-axis can be the first reference axis and the second reference axis, and Z axis is third reference axis.
Referring to fig. 2, which is a kind of coordinate frame system schematic diagram provided in an embodiment of the present invention.Interface shown in it
Interface is overlooked for two dimension, X-direction is upward, and Y direction is to the left.The dotted data of white in Fig. 2 are in laser point cloud data
Data point, arc-shaped dotted line and linear dash line in interface are auxiliary line, and the auxiliary line is for limiting mark range.
Overlooking the second sample laser point cloud data for showing in interface in two dimension can check for mark person, so as to mark to
Mark object.Mark person can be people or high-grade intelligent robot etc., and the application is not especially limited this.
Since sample laser point cloud data is distributed across the data in solid space, the mark of object to be marked can be surrounded
Frame is also solid, this needs mark person to be operated in solid space, and operation difficulty is larger.For the ease of mark person to ginseng
It examines callout box to be adjusted, this step is shown in two dimension vertical view interface by second laser point cloud data and with reference to callout box.
S160: it obtains mark person and overlooks the second sample laser point cloud data of interface display for two dimension and refer to callout box
The first adjustment of input operates, and is operated according to the first adjustment, determines the mark of object to be marked in the second sample laser point cloud data
Quasi- callout box.
Wherein, the first adjustment operation is used to adjust the reference callout box comprising object to be marked.The first adjustment operation can be with
Including at least one of mouse clicking operation, mouse drag operation, keyboard drag operation etc..
It is operated according to for the first adjustment with reference to callout box, available callout box adjusted, the mark adjusted
Note frame can be used as standard callout box.
Referring to Fig. 3, the upper right corner is that two dimension vertical view interface display is the second sample laser point cloud data, wherein the ginseng shown
Examining callout box is the two-dimensional rectangle frame in X/Y plane, and the adjustment operation of mark person can be carried out for the two-dimensional rectangle frame, adjustment
Easy implementation easy to operate.The X and Y coordinates with reference to callout box vertex can be modified according to the operation of the first adjustment of mark person, are referred to
The Z coordinate on callout box vertex remains unchanged.
S170: training set is added in the second sample laser point cloud data and corresponding standard callout box, the second sample is swashed
Light point cloud data and corresponding standard callout box are as model training data, using updated network model as target network mould
Type returns to step S120.
In the present embodiment, when training set is added in the second sample laser point cloud data, data set to be marked can also be deleted
In the second sample laser point cloud data.
Using the second sample laser point cloud data and corresponding standard callout box as model training data, by updated net
Network model returns to step S120 as target network model, can continue to be trained target network model, constantly mention
The accuracy for the reference callout box that high target network model determines, and then mark person is reduced to the adjustment amount of reference callout box, it mentions
Height determines efficiency when standard callout box.
As shown in the above, the present embodiment can mark the sample laser point cloud data marked and corresponding standard
Frame is trained target network model, is determined using the network model after training to be marked in the laser point cloud data not marked
The reference callout box of object, as reference, to be labeled for mark person with reference to callout box, obtain standard callout box, such energy
Enough mark complexities for reducing mark person, improve the annotating efficiency of training set data；Simultaneously as laser point cloud data is distribution
Data in three-dimensional space overlook in interface in two dimension and show laser point cloud data and reference callout box for mark person to reference
Callout box is adjusted, and then obtains standard callout box, three-dimensional encirclement frame is directly marked in three dimensions instead of original, two
Dimension is overlooked in interface and is adjusted to reference callout box, can largely be reduced difficulty when mark person's mark, be improved
The annotating efficiency of training set data.
In another embodiment of the invention, it is based on embodiment illustrated in fig. 1, the second sample laser is determined in step S140
In point cloud data after the reference callout box of object to be marked, this method can also show the second sample laser point cloud data
In Three Dimensional Interface.Wherein, Three Dimensional Interface is corresponding with above-mentioned three-dimensional system of coordinate.
It is operated in step S160 according to the first adjustment, determines the standard of object to be marked in the second sample laser point cloud data
The step of callout box, specifically includes following steps 1a~3a.
Step 1a: operating according to the first adjustment, determine first of object to be marked in the second sample laser point cloud data to
Adjust callout box.
In this step, operated according to for the first adjustment with reference to callout box, available callout box adjusted, the tune
Callout box after whole can be used as the first callout box to be adjusted.
Step 2a: the first callout box to be adjusted is shown in Three Dimensional Interface.
Step 3a: mark person is obtained for the second adjustment behaviour of the first callout box input to be adjusted shown in Three Dimensional Interface
Make, is operated according to second adjustment, the first callout box to be adjusted is adjusted, obtained in the second sample laser point cloud data wait mark
Infuse the standard callout box of object.
It, will due to the adjustment that the first callout box to be adjusted is the size and position progress to reference callout box in X/Y plane
After first callout box to be adjusted is shown in Three Dimensional Interface, can according to mark person input second adjustment operate, to first to
The Z parameter of adjustment callout box is adjusted, and obtains the standard callout box of object to be marked in the second sample laser point cloud data.
Second adjustment operation is used to adjust size and the position of the first callout box to be adjusted.Second adjustment operates
At least one of mouse clicking operation, mouse drag operation, keyboard drag operation etc..
In the present embodiment, when two dimension is overlooked and adjusted in interface with reference to callout box, it can be operated according to the first adjustment, it is real
When by the first adjustment operation be shown in Three Dimensional Interface.It, can be with when adjusting the first callout box to be adjusted in Three Dimensional Interface
It is operated according to second adjustment, will show for the adjustment of the first callout box to be adjusted overlooked in interface in two dimension in real time.
To sum up, the present embodiment can be overlooked in interface and Three Dimensional Interface simultaneously in two dimension and show the second sample laser point cloud number
According to the first callout box to be adjusted, three peacekeepings two dimension between mapping reference so that result data is more credible.Meanwhile also can
Enough so that mark person checks callout box, time update callout box from more perspective.
In another embodiment of the invention, it is based on embodiment illustrated in fig. 1, in order to further increase the standard of standard callout box
True property is operated according to the first adjustment in step S160, determines the standard mark of object to be marked in the second sample laser point cloud data
The step of infusing frame, can specifically include following steps 1b~5b.
Step 1b: operating according to the first adjustment, determine second of object to be marked in the second sample laser point cloud data to
Adjust callout box.
In this step, operated according to for the first adjustment with reference to callout box, available callout box adjusted, the tune
Callout box after whole can be used as the second callout box to be adjusted.
Step 2b: determine other two-dimensional coordinate systems where the other faces of the second callout box to be adjusted and three-dimensional system of coordinate it
Between the second mapping relations.
Wherein, other faces include the back side and/or side, other two-dimensional coordinate systems include: two dimension back apparent coordinates system and/or two
Tie up side view coordinate system.Other faces can be understood as in the second callout box to be adjusted in addition to two dimension overlook in interface the face that shows it
Overlook outside and with two dimension the not parallel face in the face shown in interface.The back side, refer to station the position of laser radar towards second to
The plane faced when adjusting callout box or the face parallel with the plane that this is faced.Side refers to station in the surface of position of laser radar
Second callout box to be adjusted lateral plane when to the second callout box to be adjusted, can be left side or right side.
In this step, other two-dimensional coordinates can be determined according to coordinate of second callout box to be adjusted in three-dimensional system of coordinate
The second mapping relations between system and three-dimensional system of coordinate.The second mapping relations can be determined by coordinate conversion.
Fig. 4 is a kind of schematic diagram that view, side view coordinate system were overlooked, carried on the back to three-dimensional system of coordinate provided by the invention and two dimension.Its
In, the cuboid in three-dimensional system of coordinate is the second callout box to be adjusted of object to be marked, and X-direction is the row of smart machine
Into direction, straight up, O point is the position of laser radar to Z axis, and three sides of cuboid are respectively parallel in three-dimensional system of coordinate
Three reference axis.It includes X-axis and Y-axis in coordinate system that two dimension, which is overlooked, includes Z axis and Y-axis, two-dimentional side view in two dimension back apparent coordinates system
It include Z axis and X-axis in coordinate system.It can determine that two dimension overlooks coordinate system, two dimension back apparent coordinates system, two-dimentional side view coordinate according to the figure
System's corresponding relationship between three-dimensional system of coordinate respectively.Wherein, two dimension vertical view coordinate system is corresponding with two dimension vertical view interface, two dimension back
Apparent coordinates system is corresponding with two dimension back visual interface, and two-dimentional side view coordinate system is corresponding with two-dimentional side view interface.
When the side of cuboid is not parallel to reference axis, it can be projected, be obtained according to the angle between side and reference axis
Coordinate system to where the back side and side of cube.
Step 3b: according to the second mapping relations, the second sample laser point cloud data and the second callout box to be adjusted are shown
In other two-dimentional interfaces corresponding with other two-dimensional coordinate systems.
Referring to Fig. 3, two interfaces in the lower right corner are respectively two dimension back visual interface and two-dimentional side view interface, and which show two
The the second encirclement frame to be adjusted and the second sample laser point cloud data of dimension.Each two-dimentional interface of display can be for mark person very
Second encirclement frame to be adjusted determined by checking well, to be more accurately modified to the second side to be adjusted for surrounding frame.
The lower right corner word segment of Fig. 3 also shows the button that object type to be marked can be selected for mark person.These
The corresponding object type to be marked of button options includes car, truck, bus, two wheeler, pedestrian, tricycle and not
Know.According to mark person by clicking button inputted selection operation, object type to be marked can be determined.
Step 4b: the third adjustment of second to be adjusted callout box input of the mark person for other two-dimentional interface displays is obtained
Operation.
Wherein, third adjustment operation is used to adjust size and the position of the second callout box to be adjusted.Third adjustment operation can
To include at least one of mouse clicking operation, mouse drag operation, keyboard drag operation etc..
Step 5b: the second callout box to be adjusted is adjusted according to third adjustment operation, obtains the second sample laser point
The standard callout box of object to be marked in cloud data.
After being adjusted to the second callout box to be adjusted, the standard callout box obtained after adjustment can be displayed in real time
It is overlooked in interface and Three Dimensional Interface in two dimension.
To sum up, the present embodiment swashs the second callout box to be adjusted and the second sample when obtaining the second callout box to be adjusted
Light point cloud data is shown in two dimension back view and/or side view interface, this allows mark person to check the from more kinds of views
Two callout box to be adjusted, and the second callout box to be adjusted is adjusted, in two-dimentional interface to the second callout box to be adjusted into
Row adjustment, is more easily performed, while can also improve the accuracy of standard callout box, compared to directly in three-dimensional for mark person
Other faces are adjusted in interface, annotating efficiency can be turned up.
In another embodiment of the invention, it is based on embodiment illustrated in fig. 1, when updated network model determines the second sample
When object to be marked being not present in this laser point cloud data, negative sample can also be added in the second sample laser point cloud data by this method
In this training set.
Wherein, the sample laser point cloud data refusal in negative sample training set is shown for mark person.For example, when to
It, can not when not including the laser data point of vehicle reflection in the second sample laser point cloud data when mark object is vehicle
The second sample laser point cloud data is shown that being identified again in the second sample laser point cloud data without mark person is to mark person
It is no that there are objects to be marked.Sample laser point cloud data in negative sample training set, which can be used as, instructs other network models
Negative sample when practicing.
When, there are when object to be marked, the second sample laser point cloud data can be used as in the second sample laser point cloud data
Positive sample is added in corresponding training set.
To sum up, in the present embodiment, when network model determines that there is no objects to be marked in the second sample laser point cloud data
When, the sample laser point cloud data is no longer shown to mark person and is identified, but be directly added into negative sample training set, it supplies
Other model trainings use.When there are great amount of samples data, such operation can screen sample data, will only deposit
Mark person is supplied in the sample laser point cloud data of object to be marked to be labeled, and can be improved the efficiency of mark.
In another embodiment of the invention, embodiment illustrated in fig. 1, step S120, according to above-mentioned model training number are based on
According to can specifically include following steps 1c~4c the step of being trained to target network model.
Step 1c: will be in the sample laser point cloud data input target network model in model training data.
Wherein, target network model includes feature extraction layer and recurrence layer.Sample laser point cloud in model training data
Data can be first sample laser point cloud data or the second sample laser point cloud data.
Step 2c: by the first model parameter of feature extraction layer, determining the feature vector in sample laser point cloud data,
By returning the second model parameter of layer, feature vector is returned, initial callout box is obtained.
The initial value of first model parameter and the second model parameter can rule of thumb be preset, such as can be set to
Lesser value.During each training, the first model parameter and the second model parameter are constantly corrected, and are moved closer to true
Real value.
Step 3c: initial callout box standard mark corresponding with sample laser point cloud data in above-mentioned model training data is determined
Infuse the difference between frame.Wherein, which can be obtained using loss function.
Step 4c: when above-mentioned difference is greater than default discrepancy threshold, the first model parameter and second is modified according to the difference
Model parameter returns to step 1c.When the difference is not more than default discrepancy threshold, determine that target network model training is complete
At.
When return step 1c, the sample laser point cloud data inputted in target network model is different from inputting in last time circulation
Sample laser point cloud data in target network model.
When measures of dispersion is greater than default discrepancy threshold, it is believed that the difference between the prediction result and true value of target network model
It is different larger, it needs to continue to train network.It, can when being modified according to difference to above-mentioned first model parameter and the second model parameter
The first model parameter and second are adjusted according to the specific value round about with the specific value and direction of reference difference
Model parameter.
To sum up, it present embodiments provides and target network model is carried out using sample laser point cloud data and standard callout box
The embodiment of circuit training.
In another embodiment of the invention, it is based on embodiment illustrated in fig. 1, when the sample not marked in data set to be marked
When the quantity of laser point cloud data is not more than the first quantity, that is, when the sample laser point cloud not marked in data set to be marked
When the quantity of data is not many, the sample laser point cloud data not marked can not be inputted to updated network model, but
Execute following steps 1d~4d.
Step 1d: using the sample laser point cloud data not marked in data set to be marked as third sample laser point cloud number
According to directly displaying third sample laser point cloud data.
Third sample laser point cloud data is shown specifically, can overlook in interface in two dimension.It can also be simultaneously in three-dimensional
Third sample laser point cloud data is shown in interface.
Step 2d: the labeling operation that mark person is directed to the input of third sample laser point cloud data is obtained.
Wherein, the labeling operation may include in mouse clicking operation, mouse drag operation, keyboard drag operation etc. extremely
Few one kind.
Step 3d: according to the labeling operation, the standard of the object to be marked for third sample laser point cloud data is determined
Callout box.
Step 4d: training set is added in third sample laser point cloud data and corresponding standard callout box.
To sum up, in the present embodiment, when the quantity for the sample laser point cloud data not marked in data set to be marked is not many
When, without determining the second sample laser point cloud data from the sample laser point cloud data not marked in data set to be marked,
Sample laser point cloud data without that will not mark inputs updated network model, but directly displays the sample not marked and swash
Light point cloud data determines standard callout box.
The embodiment of the present invention is illustrated in conjunction with specific example below.
Execution flow diagram shown in Figure 5, in the initial stage, sample laser point cloud data is after manually marking
Training set is added, the sample laser point cloud data marked in training set is learnt using machine learning algorithm, obtains
Base neural network model.The base can be used before artificial mark for the subsequent sample laser point cloud data not marked
Plinth neural network model detects the sample laser point cloud data that this part does not mark, filters out satisfactory data
(for example, filtering out, there are the data of object to be marked), and generate corresponding with the data filtered out with reference to callout box, the reference
Callout box may be used as the auxiliary information of subsequent artefacts' mark, to improve the efficiency of mark.Swash by the sample manually marked
Light point cloud data enters back into the process of trained neural network model, and training gained neural network model will replace old neural network
Model circuits sequentially progress, completes until all laser point cloud datas mark for screening next time.
Screening logical schematic shown in Figure 6 to sample laser point cloud data.Based on depth learning technology, using
Mark sample laser point cloud data training obtains neural network model.Screening Treatment device in electronic equipment loads neural network mould
Type, and each sample laser point cloud data that do not mark is analyzed, whether retained the selection result of the data, works as screening
As a result corresponding with reference to callout box also to be obtained when the reservation data.Screening Treatment device in electronic equipment can be one,
It can be multiple.When screening to sample laser point cloud data, each sample laser point cloud data of input is by mind
When each level through network model, this layer of meeting classifies to data according to the characteristic value of current layer.Pass through neural network mould
After several levels of type, a the selection result, the selection result can be exported to the object to be marked in sample laser point cloud data
There are the probability of object to be marked in characterization sample laser point cloud data, and the encirclement frame including that can surround object to be marked.Cause
This comes out the high data screening of probability by judging probability size, is shown to corresponding encirclement frame to mark
Note person's adjustment, and then improve the efficiency of subsequent artefacts' mark.
Fig. 7 is a kind of structural schematic diagram of the annotation equipment of training set data provided in an embodiment of the present invention.The device is real
Example is applied applied to electronic equipment.The Installation practice is corresponding with embodiment illustrated in fig. 1.The device includes:
Data acquisition module 710 is configured as obtaining in training set the first sample laser point cloud data that has marked and right
The standard callout box for the object to be marked answered, as model training data；
Model training module 720 is configured as being trained target network model according to model training data, obtaining
Updated network model；Wherein, updated network model is used for so that sample laser point cloud data and corresponding standard mark
It is interrelated to infuse frame；
Quantity judgment module 730 is configured as the sample laser point cloud data for judging not mark in data set to be marked
Whether quantity is greater than the first quantity；
Reference block determining module 740 is configured as the sample laser point cloud data not marked in data set to be marked
When quantity is greater than the first quantity, the second quantity sample is determined from the sample laser point cloud data not marked in data set to be marked
This laser point cloud data determines the second sample laser point by updated network model as the second sample laser point cloud data
The reference callout box of object to be marked in cloud data；
Two-dimentional display module 750 is configured as the first mapping overlooked between coordinate system and three-dimensional system of coordinate according to two dimension
Relationship is shown in two dimension vertical view interface by the second sample laser point cloud data and with reference to callout box；Wherein, three-dimensional system of coordinate is
Coordinate system where second sample laser point cloud data, it is corresponding with two dimension vertical view coordinate system that two dimension overlooks interface；
Template determining module 760 is configured as the second sample that acquisition mark person overlooks interface display for two dimension and swashs
Light point cloud data and the first adjustment inputted with reference to callout box operate, and are operated according to the first adjustment, determine the second sample laser point
The standard callout box of object to be marked in cloud data；
First is added module 770, is configured as the second sample laser point cloud data and corresponding standard callout box being added
Training set；
Data update module 780, be configured as using the second sample laser point cloud data and corresponding standard callout box as
Model training data return to model training module 720 using updated network model as target network model, execute basis
Model training data are trained target network model, obtain the operation of updated network model.
In another embodiment of the invention, embodiment illustrated in fig. 7, the device are based on further include:
Three-dimensional Display module (not shown), it is to be marked right in determining the second sample laser point cloud data to be configured as
After the reference callout box of elephant, the second sample laser point cloud data is shown in Three Dimensional Interface；Wherein, Three Dimensional Interface and three-dimensional
Coordinate system is corresponding；
Template determining module 760, operates according to the first adjustment, and it is to be marked right in the second sample laser point cloud data to determine
When the standard callout box of elephant, comprising:
It is operated according to the first adjustment, determines the first mark to be adjusted of object to be marked in the second sample laser point cloud data
Frame；
The first callout box to be adjusted is shown in Three Dimensional Interface；
The second adjustment that mark person is obtained for the first callout box input to be adjusted shown in Three Dimensional Interface operates, according to
Second adjustment operation, is adjusted the first callout box to be adjusted, obtains object to be marked in the second sample laser point cloud data
Standard callout box.
In another embodiment of the invention, it is based on embodiment illustrated in fig. 7, template determining module 760 is adjusted according to first
Whole operation, when determining the standard callout box of object to be marked in the second sample laser point cloud data, comprising:
It is operated according to the first adjustment, determines the second mark to be adjusted of object to be marked in the second sample laser point cloud data
Frame；
Determine between other two-dimensional coordinate systems and three-dimensional system of coordinate where the other faces of the second callout box to be adjusted
Two mapping relations；Wherein, other faces include the back side and/or side, other two-dimensional coordinate systems include: two dimension back apparent coordinates system and/
Or two-dimentional side view coordinate system；
According to the second mapping relations, by the second sample laser point cloud data and the second callout box to be adjusted show with other
In other corresponding two-dimentional interfaces of two-dimensional coordinate system；
The third for obtaining second to be adjusted callout box input of the mark person for other two-dimentional interface displays adjusts operation；
The second callout box to be adjusted is adjusted according to third adjustment operation, is obtained in the second sample laser point cloud data
The standard callout box of object to be marked.
In another embodiment of the invention, embodiment illustrated in fig. 7, device are based on further include:
Second is added module (not shown), is configured as determining the second sample laser point when updated network model
When object to be marked being not present in cloud data, the second sample laser point cloud data is added in negative sample training set；Negative sample instruction
Practice the sample laser point cloud data refusal concentrated to be shown for mark person.
In another embodiment of the invention, it is based on embodiment illustrated in fig. 7, model training module 720 is specifically configured
Are as follows:
It will be in the sample laser point cloud data input target network model in model training data；Target network model includes
Feature extraction layer and recurrence layer；
By the first model parameter of feature extraction layer, the feature vector in sample laser point cloud data is determined；By returning
The second model parameter for returning layer, returns feature vector, obtains initial callout box；
It determines between initial callout box standard callout box corresponding with sample laser point cloud data in model training data
Difference；
When difference is greater than default discrepancy threshold, the first model parameter and the second model parameter are modified according to difference, returned
Execute operation sample laser point cloud data inputted in target network model；
When difference is not more than default discrepancy threshold, determine that target network model training is completed.
In another embodiment of the invention, embodiment illustrated in fig. 7, the device are based on further include:
Two-dimentional display module 750 is additionally configured to the sample laser point cloud data not marked in data set to be marked
When quantity is not more than the first quantity, using the sample laser point cloud data not marked in data set to be marked as third sample laser
Point cloud data directly displays third sample laser point cloud data；
Template determining module 760 is additionally configured to obtain mark person for the input of third sample laser point cloud data
Labeling operation；According to labeling operation, the standard callout box of the object to be marked for third sample laser point cloud data is determined；
First is added module 770, is additionally configured to add third sample laser point cloud data and corresponding standard callout box
Enter training set.
Above-mentioned apparatus embodiment is corresponding with embodiment of the method, has same technical effect, tool with this method embodiment
Body illustrates referring to embodiment of the method.Installation practice is obtained based on embodiment of the method, and specific description may refer to method
Embodiment part, details are not described herein again.
Those of ordinary skill in the art will appreciate that: attached drawing is the schematic diagram of one embodiment, module in attached drawing or
Process is not necessarily implemented necessary to the present invention.
Those of ordinary skill in the art will appreciate that: the module in device in embodiment can describe to divide according to embodiment
It is distributed in the device of embodiment, corresponding change can also be carried out and be located in one or more devices different from the present embodiment.On
The module for stating embodiment can be merged into a module, can also be further split into multiple submodule.
Finally, it should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although
Present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: it still may be used
To modify to technical solution documented by previous embodiment or equivalent replacement of some of the technical features；And
These are modified or replaceed, the spirit and model of technical solution of the embodiment of the present invention that it does not separate the essence of the corresponding technical solution
It encloses.
Claims (10)
1. a kind of mask method of training set data characterized by comprising
The standard callout box of the first sample laser point cloud data and corresponding object to be marked that have marked in training set is obtained, is made
For model training data；
According to the model training data, target network model is trained, obtains updated network model；Wherein, institute
Updated network model is stated for so that sample laser point cloud data and corresponding standard callout box are interrelated；
Judge whether the quantity for the sample laser point cloud data not marked in data set to be marked is greater than the first quantity；
If it is greater, then determining the second quantity sample from the sample laser point cloud data not marked in the data set to be marked
This laser point cloud data determines that second sample swashs by updated network model as the second sample laser point cloud data
The reference callout box of object to be marked in light point cloud data；
The first mapping relations between coordinate system and three-dimensional system of coordinate are overlooked according to two dimension, by the second sample laser point cloud number
It is overlooked in interface according to two dimension is shown in reference callout box；Wherein, the three-dimensional system of coordinate is the second sample laser point cloud
Coordinate system where data, it is corresponding with the two dimension vertical view coordinate system that the two dimension overlooks interface；
Mark person is obtained to overlook the second sample laser point cloud data of interface display for the two dimension and input with reference to callout box
The first adjustment operation, according to the first adjustment operate, determine object to be marked in the second sample laser point cloud data
Standard callout box；
The training set is added in the second sample laser point cloud data and corresponding standard callout box；By second sample
Laser point cloud data and corresponding standard callout box are as model training data, using updated network model as target network
Model, return execution is described to be trained target network model according to the model training data, obtains updated network
The step of model.
2. the method as described in claim 1, which is characterized in that be marked in determining the second sample laser point cloud data
After the reference callout box of object, the method also includes:
The second sample laser point cloud data is shown in Three Dimensional Interface；Wherein, the Three Dimensional Interface and the three-dimensional seat
Mark system corresponds to；
It is described to be operated according to the first adjustment, determine the standard mark of object to be marked in the second sample laser point cloud data
The step of infusing frame, comprising:
It is operated according to the first adjustment, determines that first of object to be marked in the second sample laser point cloud data is to be adjusted
Callout box；
The described first callout box to be adjusted is shown in the Three Dimensional Interface；
The second adjustment that mark person is obtained for the first callout box input to be adjusted shown in the Three Dimensional Interface operates, according to
The second adjustment operation, is adjusted the described first callout box to be adjusted, obtains the second sample laser point cloud data
In object to be marked standard callout box.
3. the method as described in claim 1, which is characterized in that it is described to be operated according to the first adjustment, determine described second
In sample laser point cloud data the step of the standard callout box of object to be marked, comprising:
It is operated according to the first adjustment, determines that second of object to be marked in the second sample laser point cloud data is to be adjusted
Callout box；
It determines between other two-dimensional coordinate systems and the three-dimensional system of coordinate where the other faces of the described second callout box to be adjusted
The second mapping relations；Wherein, the other faces include the back side and/or side, other described two-dimensional coordinate systems include: two-dimentional back
Apparent coordinates system and/or two-dimentional side view coordinate system；
According to second mapping relations, the second sample laser point cloud data and second callout box to be adjusted are shown
In other two-dimentional interfaces corresponding with other described two-dimensional coordinate systems；
The third for obtaining second to be adjusted callout box input of the mark person for other two-dimentional interface displays adjusts operation；
The described second callout box to be adjusted is adjusted according to third adjustment operation, obtains the second sample laser point
The standard callout box of object to be marked in cloud data.
4. the method as described in claim 1, which is characterized in that when updated network model determines the second sample laser
When object to be marked being not present in point cloud data, the method also includes:
The second sample laser point cloud data is added in negative sample training set；Sample laser in the negative sample training set
Point cloud data refusal is shown for mark person.
5. the method as described in claim 1, which is characterized in that it is described according to the model training data, to target network mould
The step of type is trained, comprising:
It will be in the sample laser point cloud data input target network model in the model training data；The target network model
Including feature extraction layer and return layer；
By the first model parameter of the feature extraction layer, the feature vector in the sample laser point cloud data is determined；It is logical
Second model parameter for returning layer is crossed, described eigenvector is returned, initial callout box is obtained；
Determine initial callout box standard callout box corresponding with sample laser point cloud data in the model training data it
Between difference；
When the difference is greater than default discrepancy threshold, first model parameter and second mould are modified according to the difference
Shape parameter returns and executes the step inputted the sample laser point cloud data in target network model；
When the difference is not more than default discrepancy threshold, determine that the target network model training is completed.
6. method as claimed in any one of claims 1 to 5, which is characterized in that when what is do not marked in the data set to be marked
When the quantity of sample laser point cloud data is not more than the first quantity, the method also includes:
Using the sample laser point cloud data not marked in the data set to be marked as third sample laser point cloud data, directly
Show the third sample laser point cloud data；
Obtain the labeling operation that mark person is directed to third sample laser point cloud data input；
According to the labeling operation, the standard mark for the object to be marked of the third sample laser point cloud data is determined
Frame；
The training set is added in the third sample laser point cloud data and corresponding standard callout box.
7. a kind of annotation equipment of training set data characterized by comprising
Data acquisition module is configured as obtaining in training set the first sample laser point cloud data that has marked and corresponding wait mark
The standard callout box for infusing object, as model training data；
Model training module is configured as being trained target network model according to the model training data, being updated
Network model afterwards；Wherein, the updated network model is used for so that sample laser point cloud data and corresponding standard mark
It is interrelated to infuse frame；
Quantity judgment module, be configured as the sample laser point cloud data for judging not mark in data set to be marked quantity whether
Greater than the first quantity；
Reference block determining module is configured as the quantity for the sample laser point cloud data not marked in the data set to be marked
When greater than the first quantity, the second quantity sample is determined from the sample laser point cloud data not marked in the data set to be marked
This laser point cloud data determines that second sample swashs by updated network model as the second sample laser point cloud data
The reference callout box of object to be marked in light point cloud data；
Two-dimentional display module is configured as the first mapping relations overlooked between coordinate system and three-dimensional system of coordinate according to two dimension, will
The second sample laser point cloud data and reference callout box are shown in two dimension and overlook in interface；Wherein, the three-dimensional system of coordinate
For the coordinate system where the second sample laser point cloud data, the two dimension overlooks interface and the two dimension overlooks coordinate system pair
It answers；
Template determining module is configured as obtaining the second sample laser point that mark person overlooks interface display for the two dimension
Cloud data and the first adjustment inputted with reference to callout box operate, and are operated according to the first adjustment, determine that second sample swashs
The standard callout box of object to be marked in light point cloud data；
First is added module, and being configured as will be described in the second sample laser point cloud data and the addition of corresponding standard callout box
Training set；
Data update module is configured as using the second sample laser point cloud data and corresponding standard callout box as model
Training data, using updated network model as target network model, return execute it is described according to the model training data,
Target network model is trained, the operation of updated network model is obtained.
8. device as claimed in claim 7, which is characterized in that described device further include:
Three-dimensional Display module is configured as the reference mark of the object to be marked in determining the second sample laser point cloud data
After frame, the second sample laser point cloud data is shown in Three Dimensional Interface；Wherein, the Three Dimensional Interface and the three-dimensional
Coordinate system is corresponding；
The template determining module, according to the first adjustment operate, determine in the second sample laser point cloud data to
When marking the standard callout box of object, comprising:
It is operated according to the first adjustment, determines that first of object to be marked in the second sample laser point cloud data is to be adjusted
Callout box；
The described first callout box to be adjusted is shown in the Three Dimensional Interface；
The second adjustment that mark person is obtained for the first callout box input to be adjusted shown in the Three Dimensional Interface operates, according to
The second adjustment operation, is adjusted the described first callout box to be adjusted, obtains the second sample laser point cloud data
In object to be marked standard callout box.
9. device as claimed in claim 7, which is characterized in that the template determining module is grasped according to the first adjustment
Make, when determining the standard callout box of object to be marked in the second sample laser point cloud data, comprising:
It is operated according to the first adjustment, determines that second of object to be marked in the second sample laser point cloud data is to be adjusted
Callout box；
It determines between other two-dimensional coordinate systems and the three-dimensional system of coordinate where the other faces of the described second callout box to be adjusted
The second mapping relations；Wherein, the other faces include the back side and/or side, other described two-dimensional coordinate systems include: two-dimentional back
Apparent coordinates system and/or two-dimentional side view coordinate system；
According to second mapping relations, the second sample laser point cloud data and second callout box to be adjusted are shown
In other two-dimentional interfaces corresponding with other described two-dimensional coordinate systems；
The third for obtaining second to be adjusted callout box input of the mark person for other two-dimentional interface displays adjusts operation；
The described second callout box to be adjusted is adjusted according to third adjustment operation, obtains the second sample laser point
The standard callout box of object to be marked in cloud data.
10. device as claimed in claim 7, which is characterized in that described device further include:
Second is added module, is configured as not depositing in the second sample laser point cloud data when updated network model determines
In object to be marked, the second sample laser point cloud data is added in negative sample training set；The negative sample training set
In sample laser point cloud data refusal shown for mark person.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910443047.8A CN110176078B (en) | 2019-05-26 | 2019-05-26 | Method and device for labeling training set data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910443047.8A CN110176078B (en) | 2019-05-26 | 2019-05-26 | Method and device for labeling training set data |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110176078A true CN110176078A (en) | 2019-08-27 |
| CN110176078B CN110176078B (en) | 2022-06-10 |
Family
ID=67695751
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910443047.8A Active CN110176078B (en) | 2019-05-26 | 2019-05-26 | Method and device for labeling training set data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110176078B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110782517A (en) * | 2019-10-10 | 2020-02-11 | 北京地平线机器人技术研发有限公司 | Point cloud marking method and device, storage medium and electronic equipment |
| CN110865756A (en) * | 2019-11-12 | 2020-03-06 | 苏州智加科技有限公司 | Image labeling method, device, equipment and storage medium |
| CN110889463A (en) * | 2019-12-10 | 2020-03-17 | 北京奇艺世纪科技有限公司 | Sample labeling method and device, server and machine-readable storage medium |
| CN111401321A (en) * | 2020-04-17 | 2020-07-10 | Oppo广东移动通信有限公司 | Object recognition model training method and device, electronic equipment and readable storage medium |
| CN111598006A (en) * | 2020-05-18 | 2020-08-28 | 北京百度网讯科技有限公司 | Method and device for labeling objects |
| CN112200274A (en) * | 2020-12-09 | 2021-01-08 | 湖南索莱智能科技有限公司 | Target detection method and device, electronic equipment and storage medium |
| CN112669373A (en) * | 2020-12-24 | 2021-04-16 | 北京亮道智能汽车技术有限公司 | Automatic labeling method and device, electronic equipment and storage medium |
| CN112862016A (en) * | 2021-04-01 | 2021-05-28 | 北京百度网讯科技有限公司 | Method, device and equipment for labeling objects in point cloud and storage medium |
| CN113592897A (en) * | 2020-04-30 | 2021-11-02 | 初速度（苏州）科技有限公司 | Point cloud data labeling method and device |
| CN113673622A (en) * | 2021-08-31 | 2021-11-19 | 三一专用汽车有限责任公司 | Laser point cloud data labeling method, device, equipment and product |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107093210A (en) * | 2017-04-20 | 2017-08-25 | 北京图森未来科技有限公司 | A kind of laser point cloud mask method and device |
| CN107784038A (en) * | 2016-08-31 | 2018-03-09 | 法乐第(北京)网络科技有限公司 | A kind of mask method of sensing data |
| CN108154560A (en) * | 2018-01-25 | 2018-06-12 | 北京小马慧行科技有限公司 | Laser point cloud mask method, device and readable storage medium storing program for executing |
-
2019
- 2019-05-26 CN CN201910443047.8A patent/CN110176078B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107784038A (en) * | 2016-08-31 | 2018-03-09 | 法乐第(北京)网络科技有限公司 | A kind of mask method of sensing data |
| CN107093210A (en) * | 2017-04-20 | 2017-08-25 | 北京图森未来科技有限公司 | A kind of laser point cloud mask method and device |
| CN108154560A (en) * | 2018-01-25 | 2018-06-12 | 北京小马慧行科技有限公司 | Laser point cloud mask method, device and readable storage medium storing program for executing |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110782517A (en) * | 2019-10-10 | 2020-02-11 | 北京地平线机器人技术研发有限公司 | Point cloud marking method and device, storage medium and electronic equipment |
| CN110865756A (en) * | 2019-11-12 | 2020-03-06 | 苏州智加科技有限公司 | Image labeling method, device, equipment and storage medium |
| CN110889463A (en) * | 2019-12-10 | 2020-03-17 | 北京奇艺世纪科技有限公司 | Sample labeling method and device, server and machine-readable storage medium |
| CN111401321A (en) * | 2020-04-17 | 2020-07-10 | Oppo广东移动通信有限公司 | Object recognition model training method and device, electronic equipment and readable storage medium |
| CN113592897B (en) * | 2020-04-30 | 2024-03-29 | 魔门塔(苏州)科技有限公司 | Point cloud data labeling method and device |
| CN113592897A (en) * | 2020-04-30 | 2021-11-02 | 初速度（苏州）科技有限公司 | Point cloud data labeling method and device |
| CN111598006B (en) * | 2020-05-18 | 2023-05-26 | 阿波罗智联(北京)科技有限公司 | Method and device for labeling objects |
| CN111598006A (en) * | 2020-05-18 | 2020-08-28 | 北京百度网讯科技有限公司 | Method and device for labeling objects |
| CN112200274A (en) * | 2020-12-09 | 2021-01-08 | 湖南索莱智能科技有限公司 | Target detection method and device, electronic equipment and storage medium |
| CN112669373A (en) * | 2020-12-24 | 2021-04-16 | 北京亮道智能汽车技术有限公司 | Automatic labeling method and device, electronic equipment and storage medium |
| CN112669373B (en) * | 2020-12-24 | 2023-12-05 | 北京亮道智能汽车技术有限公司 | Automatic labeling method and device, electronic equipment and storage medium |
| CN112862016A (en) * | 2021-04-01 | 2021-05-28 | 北京百度网讯科技有限公司 | Method, device and equipment for labeling objects in point cloud and storage medium |
| CN113673622A (en) * | 2021-08-31 | 2021-11-19 | 三一专用汽车有限责任公司 | Laser point cloud data labeling method, device, equipment and product |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110176078B (en) | 2022-06-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110176078A (en) | A kind of mask method and device of training set data | |
| EP3161414B1 (en) | Locating system having a hand-held locating unit | |
| CN110136273A (en) | A kind of sample data mask method and device in machine learning | |
| CN110135453A (en) | A kind of laser point cloud data mask method and device | |
| CN108985230A (en) | Method for detecting lane lines, device and computer readable storage medium | |
| CN108764187A (en) | Extract method, apparatus, equipment, storage medium and the acquisition entity of lane line | |
| CN109870983A (en) | Handle the method, apparatus of pallet stacking image and the system for picking of storing in a warehouse | |
| CN106969763A (en) | For the method and apparatus for the yaw angle for determining automatic driving vehicle | |
| CN110281231B (en) | Three-dimensional vision grabbing method for mobile robot for unmanned FDM additive manufacturing | |
| CN107481292A (en) | The attitude error method of estimation and device of vehicle-mounted camera | |
| CN113378760A (en) | Training target detection model and method and device for detecting target | |
| CN106560835A (en) | Guideboard identification method and device | |
| CN111178170B (en) | Gesture recognition method and electronic equipment | |
| CN114841944B (en) | Tailing dam surface deformation inspection method based on rail-mounted robot | |
| CN108256454A (en) | A kind of training method based on CNN models, human face posture estimating and measuring method and device | |
| CN109583312A (en) | Lane detection method, apparatus, equipment and storage medium | |
| CN111496786A (en) | Point cloud model-based mechanical arm operation processing track planning method | |
| CN110110678A (en) | Determination method and apparatus, storage medium and the electronic device of road boundary | |
| CN115115768A (en) | Object coordinate recognition system, method, device and medium based on stereoscopic vision | |
| CN115424265A (en) | Point cloud semantic segmentation and labeling method and system | |
| CN116883612A (en) | Three-dimensional scene model generation method and system | |
| CN108664860A (en) | The recognition methods of room floor plan and device | |
| CN117115063A (en) | Multi-source data fusion application method | |
| CN114266879A (en) | Three-dimensional data enhancement method, model training detection method, three-dimensional data enhancement equipment and automatic driving vehicle | |
| CN109657540A (en) | Withered tree localization method and system |
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 | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20211130 Address after: 215100 floor 23, Tiancheng Times Business Plaza, No. 58, qinglonggang Road, high speed rail new town, Xiangcheng District, Suzhou, Jiangsu Province Applicant after: MOMENTA (SUZHOU) TECHNOLOGY Co.,Ltd. Address before: Room 601-a32, Tiancheng information building, No. 88, South Tiancheng Road, high speed rail new town, Xiangcheng District, Suzhou City, Jiangsu Province Applicant before: MOMENTA (SUZHOU) TECHNOLOGY Co.,Ltd. |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |