CN107480720B - Human body posture model training method and device - Google Patents
Human body posture model training method and device Download PDFInfo
- Publication number
- CN107480720B CN107480720B CN201710712291.0A CN201710712291A CN107480720B CN 107480720 B CN107480720 B CN 107480720B CN 201710712291 A CN201710712291 A CN 201710712291A CN 107480720 B CN107480720 B CN 107480720B
- Authority
- CN
- China
- Prior art keywords
- training
- human body
- training model
- model
- heat
- 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.)
- Active
Links
Images
Classifications
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a human body posture model training method and device, and relates to the field of computer vision. The human body posture model training method comprises the following steps: obtaining a target human body image and a first hotspot graph corresponding to the target human body image; sampling a target human body image to obtain a training sample; establishing a training model; and performing regression on the training samples according to the training model to obtain a second heat point diagram, and optimizing the training model according to the first heat point diagram and the second heat point diagram to obtain the target training model. The human body posture model training method and the human body posture model training device can improve the accuracy of a training model, reduce the volume of the training model and improve the running speed.
Description
Technical Field
The invention relates to the field of computer vision, in particular to a human body posture model training method and device.
Background
Human body posture estimation is an increasingly active research field of computer vision and has wide application prospect. The method is commonly used for man-machine interaction, intelligent monitoring, athlete auxiliary training, video coding and the like. Driven by these applications, behavioral analysis has become a research hotspot in related fields of computer vision, machine learning, pattern recognition, data mining, cognitive psychology, and the like, for the last decade.
However, most of the current research results have the defect of poor accuracy in the aspect of human posture estimation. Therefore, it has become necessary to provide a new human body posture model training method to improve the accuracy of human body posture estimation.
Disclosure of Invention
In view of the above, an object of the embodiments of the present invention is to provide a method and an apparatus for training a human body posture model to improve the above-mentioned problems.
In a first aspect, an embodiment of the present invention provides a human body posture model training method, where the method includes:
obtaining a target human body image and a first heat point diagram corresponding to the target human body image;
sampling the target human body image to obtain a training sample;
establishing a training model;
and performing regression on the training samples according to the training model to obtain a second heat point diagram, and optimizing the training model according to the first heat point diagram and the second heat point diagram to obtain a target training model.
The human body posture model training method as described above, preferably, before the obtaining of the target training model, the method further includes:
simplifying the optimized tensors in the training model;
the step of simplifying the tensor in the optimized training model comprises the following steps:
decomposing the optimized tensor in the training model;
simplifying the decomposed tensors in the training model;
and loading the simplified training model.
In the above human body posture model training method, it is preferable that the step of optimizing the training model based on the first heat map and the second heat map includes:
and optimizing the training model according to the first heat point diagram and the second heat point diagram by adopting a back propagation method and a random gradient descent algorithm.
In the above human body posture model training method, preferably, the step of obtaining the target human body image and the first heat point map corresponding to the target human body image includes:
extracting the human body image marked with the joint points in the original image to obtain the target human body image;
and converting the target human body image into the first hot spot image with the joint point as a reference and with a Gaussian distribution of heat value.
In the above human body posture model training method, preferably, the step of obtaining a second heat point map by performing regression on the training samples according to the training model includes:
and quantizing the RGB value of each pixel point in the training sample and then using the quantized RGB value as the input of the training model, using the obtained output value as the corresponding heat value of each pixel point, and generating the second heat point map according to the corresponding heat value of each pixel point.
In a second aspect, an embodiment of the present invention provides an apparatus for training a human body posture model, where the apparatus includes:
the acquisition module is used for acquiring a target human body image and a first hotspot graph corresponding to the target human body image;
the sampling module is used for sampling the target human body image to obtain a training sample;
the modeling module is used for establishing a training model;
the regression module is used for performing regression on the training sample according to the training model to obtain a second heat point diagram;
and the optimization module is used for optimizing the training model according to the first heat point diagram and the second heat point diagram to obtain a target training model.
The human body posture model training device as described above preferably further comprises:
a simplification module, configured to simplify tensors in the optimized training model;
the simplified module includes:
the decomposition submodule is used for decomposing the optimized tensor in the training model;
the simplification submodule is used for simplifying the decomposed tensors in the training model;
and the loading submodule is used for loading the simplified training model.
In the human body posture model training device, preferably, the optimization module is configured to optimize the training model according to the first heat point diagram and the second heat point diagram by using a back propagation method and a stochastic gradient descent algorithm.
The human body posture model training device as described above, preferably, the obtaining module includes:
the extraction submodule is used for extracting the human body image marked with the joint points in the original image to obtain the target human body image;
and the conversion submodule is used for converting the target human body image into the first hot spot image with the joint point as a reference and with the heat value in Gaussian distribution.
In the human body posture model training device, preferably, the regression module is configured to quantize an RGB value of each pixel point in the training sample and use the quantized RGB value as an input of the training model, obtain an output value as a heat value corresponding to each pixel point, and generate the second heat point map according to the heat value corresponding to each pixel point.
Compared with the prior art, the human body posture model training method and the device provided by the invention have the following beneficial effects:
according to the human body posture model training method and device, the training samples are regressed according to the training model to obtain the second heat point diagram, the training model is optimized according to the second heat point diagram and the first heat point diagram corresponding to the target human body image to obtain the target training model, and compared with the traditional training method, the accuracy rate of the training model is higher.
Furthermore, tensors in the training model are simplified, so that the training model is smaller in size and higher in speed.
In order to make the aforementioned and other objects, features and advantages of the present invention comprehensible, preferred embodiments accompanied with figures are described in detail below.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the embodiments will be briefly described below, it should be understood that the following drawings only illustrate some embodiments of the present invention and therefore should not be considered as limiting the scope, and for those skilled in the art, other related drawings can be obtained according to the drawings without inventive efforts.
Fig. 1 is a flowchart of a human body posture model training method according to a first embodiment of the present invention.
Fig. 2 is a flowchart of a human body posture model training method according to a second embodiment of the present invention.
Fig. 3 is a block diagram of a human body posture model training device according to a third embodiment of the present invention.
Fig. 4 is a block diagram of an acquisition module according to a third embodiment of the present invention.
Fig. 5 is a block diagram of a simplified module according to a third embodiment of the present invention.
Icon: 10-a human body posture model training device; 110-an obtaining module; 111-extraction submodule; 112-a conversion submodule; 120-a sampling module; 130-a modeling module; 140-a regression module; 150-an optimization module; 160-simplified module; 161-decomposition submodule; 162-simplified sub-module; 163-load the submodule.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. The components of embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present invention, presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present invention without making any creative effort, shall fall within the protection scope of the present invention.
It should be noted that: like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined and explained in subsequent figures. Meanwhile, in the description of the present invention, the terms "first", "second", and the like are used only for distinguishing the description, and are not to be construed as indicating or implying relative importance.
First embodiment
Referring to fig. 1, a flowchart of a human body posture model training method according to a preferred embodiment of the present invention is shown, and the specific process shown in fig. 1 will be described in detail below.
Step S101, obtaining a target human body image and a first heat point diagram corresponding to the target human body image.
In the embodiment of the invention, the human body posture model training method can be applied to a server or a user terminal, and the server can be a network server, a database server and the like. The user terminal may be a Personal Computer (PC), a tablet PC, a smart phone, a Personal Digital Assistant (PDA), or the like.
Taking application to a user terminal as an example, the user terminal stores an original image for human body posture model training in advance, and the original image includes at least one human body image. When a human body posture model needs to be established, a user can label joint points of a human body image in an original image of the user terminal, and the number of the joint points can be one or more. Preferably, in the embodiment of the present invention, the number of the joint points is plural.
After the human body image in the original image is subjected to joint point labeling, the user terminal can extract the human body image of the joint point related to the original image labeling through image contour recognition to obtain a target human body image. The RGB values of the regions except the target human body image can be reset to zero, namely, the regions except the target human body image are all converted into black, and only the target human body image is displayed at the moment.
After the target human body image is extracted, the user terminal converts the target human body image into a first heat point diagram which takes the joint point as a reference and has Gaussian distribution of heat values. Therefore, the heat value of the pixel point in the first heat point diagram is gradually reduced along the direction far away from the joint point by taking the joint point as a reference. The heat value of the pixel point represents the attention degree of the pixel point, the larger the heat value is, the higher the attention degree is represented, and the darker the color displayed on the heat point diagram is.
It should be noted that the joint point may be one pixel point or a plurality of adjacent pixel points in the target human body image. In the embodiment of the invention, the joint points are a plurality of adjacent pixel points in the target human body image.
And S102, sampling the target human body image to obtain a training sample.
After obtaining the target human body image and the first hotspot graph corresponding to the target human body image, the user terminal samples the target human body image, selects an image in a certain area in the target human body image, and takes the image in the selected area as a training sample.
The sampling of the target human body image may be performed by the user terminal according to a selection operation of a user, or may be performed by the user terminal randomly sampling the target human body image, or may be performed by the user terminal in an area where joint points in the target human body image are dense. In the embodiment of the invention, the user terminal randomly samples the target human body image.
Furthermore, in order to prevent overfitting, in the embodiment of the present invention, the sampled image may be processed, such as adding noise, rotating the image, and the like, to obtain a plurality of images, and the plurality of images may be used as training samples.
And step S103, establishing a training model.
Meanwhile, the user terminal constructs a training model for training the human body posture model, and the training model is used for returning the training sample to the heat point diagram.
The training model comprises a plurality of convolutional neural network regressors with two branches, wherein one branch of the convolutional neural network regressors is used for regressing the joint points in the training sample, and the other branch of the convolutional neural network regressors is used for regressing the regions except the joint points in the training sample.
Preferably, in the embodiment of the present invention, the training model includes 2 to 6 cascaded convolutional neural network regressors with two branches, so as to ensure the accuracy of the training model while ensuring that the volume of the training model is not too large.
In the embodiment of the present invention, the order of step S102 and step S103 is not limited.
And step S104, performing regression on the training samples according to the training model to obtain a second heat point diagram.
After the training model is established, the user terminal quantizes the RGB value of each pixel point in the training sample into a tensor corresponding to the RGB value, the quantized tensor is used as the input of the training model for operation, the output value corresponding to each pixel point is obtained, the output value corresponding to each pixel point is used as the heat value of the pixel point, and then a second heat point diagram is generated according to the heat value corresponding to each pixel point.
For example, the RGB value of a certain pixel is (100, 92, 58), that is, the luminance values of red, green and blue of the pixel are 100, 92 and 58, respectively, and the user terminal can quantize the RGB value of the pixel into a three-dimensional tensor (10, 5, 6) according to a rounding method. It is to be understood that the above quantization of the RGB values of the pixel points is merely an example, and is not a specific limitation to the quantization of the RGB values of the pixel points.
And step S105, optimizing the training model according to the first heat point diagram and the second heat point diagram.
After the training sample is returned to the second heat point diagram through the training model, the user terminal adjusts the training model according to the difference between the first heat point diagram and the second heat point diagram so as to optimize the training model and further improve the accuracy of the training model.
In the embodiment of the invention, a training model is optimized according to the first heat point diagram and the second heat point diagram by adopting a back propagation method and a random gradient descent algorithm. Specifically, after obtaining the second hotspot graph, the user terminal calculates the euclidean distance (difference between the heat values) of the corresponding pixel points on the first hotspot graph and the second hotspot graph. And the Euclidean distance of each pixel point is used as a gradient to be reversely propagated from the bottom layer to the top layer of the training model, and the training model is adjusted according to the result of the reverse propagation. And then, repeatedly executing the steps until the set iteration times are finished, and taking the finally obtained training model as a target training model.
In summary, in the human body posture model training method provided in the embodiment of the present invention, joint points of the human body image in the original image are labeled to obtain the target human body image and the first heat point map corresponding to the target human body image, the target human body image is sampled to obtain the training sample, the training sample is regressed according to the training model to obtain the second heat point map, and then the training model is iteratively optimized by using a back propagation method and a stochastic gradient descent algorithm to improve the accuracy of the training model. Compared with the traditional training method, the human body posture model training method provided by the embodiment of the invention has the advantages that the accuracy of the training model obtained by training is higher, and the human body posture estimation error is lower.
Second embodiment
Referring to fig. 2, the human body posture model training method provided by the embodiment of the present invention is a further improvement based on the first embodiment, and the parts not mentioned in the embodiment refer to the description in the first embodiment. The specific process shown in fig. 2 will be described in detail below.
In step S201, a target human body image and a first hotspot graph corresponding to the target human body image are obtained.
Step S202, sampling is carried out on the target human body image, and a training sample is obtained.
Step S203, establishing a training model.
And step S204, performing regression on the training samples according to the training model to obtain a second heat point diagram.
And S205, optimizing the training model according to the first heat point diagram and the second heat point diagram.
And step S206, simplifying tensors in the optimized training model.
The training model obtained by optimizing the training model has a large volume, a large amount of calculation when the human posture is calculated, and a low calculation speed. Therefore, in the human body posture model training method provided by the embodiment of the invention, after the training model is optimized, the tensor in the training model is simplified, so that the volume of the training model is reduced, and the calculation speed is increased.
Firstly, after optimizing the model, the user terminal decomposes the tensor in the optimized training model. The decomposed tensor is then simplified to reduce the volume of the training model. And finally, loading the simplified training model.
Preferably, in the embodiment of the present invention, the number of times of simplifying the tensor in the training model is multiple times. Specifically, after the simplified training model is loaded each time, the tensor in the training model is decomposed again to be simplified again until the set iteration times are completed, the optimal training model is finally obtained, and the finally obtained training model is used as the target training model.
In summary, in the human body posture model training method provided in the embodiment of the present invention, joint points of a human body image in an original image are labeled to obtain a target human body image and a first heat point diagram corresponding to the target human body image, the target human body image is sampled to obtain a training sample, the training sample is regressed according to the training model to obtain a second heat point diagram, then a back propagation method and a stochastic gradient descent algorithm are used to perform iterative optimization on the training model to improve the accuracy of the training model, and finally, the optimized training model is iteratively simplified to reduce the volume of the training model and improve the operation speed of the training model. Compared with the traditional training method, the human body posture model training method provided by the embodiment of the invention has the advantages that the accuracy of the training model obtained by training is higher, the training model is smaller in size, the calculation amount in human body posture estimation is greatly reduced, the human body posture estimation error is lower, and the speed is higher.
Third embodiment
Referring to fig. 3, which is a block diagram illustrating a human body posture model training apparatus 10 according to a preferred embodiment of the present invention, the human body posture model training apparatus 10 includes an obtaining module 110, a sampling module 120, a modeling module 130, a regression module 140, an optimizing module 150, and a simplifying module 160.
The obtaining module 110 is configured to obtain a target human body image and a first hotspot graph corresponding to the target human body image.
It is understood that the obtaining module 110 can be configured to perform the steps S101 and S201.
Referring to fig. 4, the obtaining module 110 includes an extracting sub-module 111 and a converting sub-module 112.
The extraction submodule 111 is configured to extract a human body image labeled with a joint point in the original image, so as to obtain a target human body image.
In the embodiment of the present invention, the human body posture model training apparatus 10 may be applied to a server or a user terminal. Taking application to a user terminal as an example, the user terminal stores an original image used for establishing a human body posture model in advance, and the original image comprises at least one human body image. When a human body posture model needs to be established, a user can label joint points of a human body image in an original image of the user terminal.
After the human body image in the original image is labeled with the joint points, the user terminal can extract the human body image labeled with the relevant joint points of the original image by using the extraction submodule 111 through image contour recognition to obtain a target human body image.
It is understood that the extraction sub-module 111 may be configured to perform the process of obtaining the target human body image in the above steps S101 and S201.
The conversion sub-module 112 is configured to convert the target human body image into a first hot spot image with a gaussian distribution of heat values based on the joint point.
After the target human body image is extracted, the user terminal converts the target human body image into a first hotspot graph with a gaussian distribution of the hotspot value by using the articulation point as a reference through the conversion sub-module 112.
It is understood that the conversion sub-module 112 may be configured to perform the process of obtaining the first heat point map corresponding to the target human body image in the above steps S101 and S201.
The sampling module 120 is configured to sample a target human body image to obtain a training sample.
After obtaining the target human body image and the first hotspot graph corresponding to the target human body image, the user terminal samples the target human body image through the sampling module 120, selects an image in a certain region of the target human body image labeled with a joint point, and takes the image in the selected region as a training sample.
It is understood that the sampling module 120 can be used to perform the steps S102 and S202.
The modeling module 130 is used to build a training model.
Meanwhile, the user terminal constructs a training model for human body posture model training through the modeling module 130, for returning the training sample to the heat point map.
The training model comprises a plurality of convolutional neural network regressors with two branches, wherein one branch of the convolutional neural network regressors is used for regressing the joint points in the training sample, and the other branch of the convolutional neural network regressors is used for regressing the regions except the joint points in the training sample.
It is understood that the modeling module 130 may be configured to perform the steps S103 and S203 described above.
The regression module 140 is configured to perform regression on the training samples according to the training model to obtain a second heat point diagram.
After the training model is established, the user terminal quantizes the RGB value of each pixel point in the training sample into a tensor corresponding to the RGB value through the regression module 140, performs operation by using the quantized tensor as the input of the training model to obtain an output value corresponding to each pixel point, uses the output value corresponding to each pixel point as the heat value thereof, and then generates a second heat point map according to the heat value corresponding to each pixel point.
It is understood that the regression module 140 may be used to perform the above steps S104 and S204.
The optimization module 150 is configured to optimize the training model according to the first heat map and the second heat map.
After the training samples are returned to the second hotspot graph by the training model, the user terminal optimizes the training model according to the first hotspot graph and the second hotspot graph by using a back propagation method and a random gradient descent algorithm through the optimization module 150.
It is understood that the optimization module 150 may be configured to perform the steps S105 and S205 described above.
The simplification module 160 is used to simplify the tensors in the optimized training model.
Referring to FIG. 5, the simplified module 160 includes a decomposition sub-module 161, a simplified sub-module 162, and a load sub-module 163.
The decomposition submodule 161 is used to decompose the tensors in the optimized training model.
The simplification submodule 162 is used to simplify the tensors in the decomposed training model.
The loading submodule 163 is used to load the simplified training model.
First, after optimizing the model, the user terminal decomposes the tensor in the optimized training model through the decomposition submodule 161. The decomposed tensor is then simplified by the simplification submodule 162 to reduce the volume of the training model. Finally, the simplified training model is loaded through the loading submodule 163.
It is understood that the simplified module 160 can be used to execute the step S206.
In summary, the human body posture model training device 10 provided in the embodiment of the present invention obtains the target human body image and the first heat point map corresponding to the target human body image by performing joint point labeling on the human body image in the original image, samples the target human body image to obtain the training sample, regresses the training sample according to the training model to obtain the second heat point map, then performs iterative optimization on the training model by using a back propagation method and a stochastic gradient descent algorithm to improve the accuracy of the training model, and finally performs iterative simplification on the optimized training model to reduce the volume of the training model and improve the operation speed of the training model. Compared with the traditional training method, the human body posture model training method provided by the embodiment of the invention has the advantages that the accuracy of the training model obtained by training is higher, the training model is smaller in size, the calculation amount in human body posture estimation is greatly reduced, the human body posture estimation error is lower, and the speed is higher.
In the embodiments provided in the present application, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative, and for example, the flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
In addition, the functional modules in the embodiments of the present invention may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.
The functions, if implemented in the form of software functional modules and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes. It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention. It should be noted that: like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined and explained in subsequent figures.
The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (6)
1. A human body posture model training method is characterized by comprising the following steps:
extracting the human body image marked with the joint points in the original image to obtain a target human body image;
converting the target human body image into a first hot spot image with the joint point as a reference and with a Gaussian distribution of heat value;
sampling the target human body image to obtain a training sample;
establishing a training model, wherein the training model comprises 2-6 cascaded convolutional neural network regressors with two branches, one branch of the convolutional neural network regressor is used for regressing the joint points in the training sample, and the other branch of the convolutional neural network regressor is used for regressing the regions except the joint points in the training sample;
and quantizing the RGB value of each pixel point in the training sample and then using the quantized RGB value as the input of the training model, using the obtained output value as the corresponding heat value of each pixel point, generating a second heat point diagram according to the corresponding heat value of each pixel point, and optimizing the training model according to the first heat point diagram and the second heat point diagram to obtain the target training model.
2. The human pose model training method of claim 1, wherein prior to the obtaining the target training model, the method further comprises:
simplifying the optimized tensors in the training model;
the step of simplifying the tensor in the optimized training model comprises the following steps:
decomposing the optimized tensor in the training model;
simplifying the decomposed tensors in the training model;
and loading the simplified training model.
3. The human pose model training method of claim 1, wherein the step of optimizing the training model according to the first and second heat point maps comprises:
and optimizing the training model according to the first heat point diagram and the second heat point diagram by adopting a back propagation method and a random gradient descent algorithm.
4. A mannequin training apparatus, the apparatus comprising:
the acquisition module is used for extracting the human body image marked with the joint points in the original image to obtain a target human body image; converting the target human body image into a first hot spot image with the joint point as a reference and with a Gaussian distribution of heat value;
the sampling module is used for sampling the target human body image to obtain a training sample;
the modeling module is used for establishing a training model, wherein the training model comprises 2-6 cascaded convolutional neural network regressors with two branches, one branch of the convolutional neural network regressors is used for regressing the joint points in the training sample, and the other branch of the convolutional neural network regressors is used for regressing the regions except the joint points in the training sample;
a regression module to: quantizing the RGB value of each pixel point in the training sample and then using the quantized RGB value as the input of the training model, using the obtained output value as the corresponding heat value of each pixel point, and generating a second heat point diagram according to the corresponding heat value of each pixel point;
and the optimization module is used for optimizing the training model according to the first heat point diagram and the second heat point diagram to obtain a target training model.
5. The mannequin training device of claim 4, further comprising:
a simplification module, configured to simplify tensors in the optimized training model;
the simplified module includes:
the decomposition submodule is used for decomposing the optimized tensor in the training model;
the simplification submodule is used for simplifying the decomposed tensors in the training model;
and the loading submodule is used for loading the simplified training model.
6. The human pose model training device of claim 4, wherein the optimization module is configured to optimize the training model based on the first heat point map and the second heat point map using a back propagation method and a stochastic gradient descent algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710712291.0A CN107480720B (en) | 2017-08-18 | 2017-08-18 | Human body posture model training method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710712291.0A CN107480720B (en) | 2017-08-18 | 2017-08-18 | Human body posture model training method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107480720A CN107480720A (en) | 2017-12-15 |
| CN107480720B true CN107480720B (en) | 2021-02-26 |
Family
ID=60600970
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710712291.0A Active CN107480720B (en) | 2017-08-18 | 2017-08-18 | Human body posture model training method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107480720B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108509838B (en) * | 2018-01-30 | 2022-03-25 | 中山大学 | Method for analyzing group dressing under joint condition |
| CN108537156B (en) * | 2018-03-30 | 2021-12-21 | 广州幻境科技有限公司 | Anti-shielding hand key node tracking method |
| CN108520223B (en) * | 2018-04-02 | 2021-11-12 | 广州方硅信息技术有限公司 | Video image segmentation method, segmentation device, storage medium and terminal equipment |
| CN111539349A (en) * | 2020-04-27 | 2020-08-14 | 平安国际智慧城市科技股份有限公司 | Training method and device of gesture recognition model, gesture recognition method and device thereof |
| CN112907583B (en) * | 2021-03-29 | 2023-04-07 | 苏州科达科技股份有限公司 | Target object posture selection method, image scoring method and model training method |
| CN116758380B (en) * | 2023-08-15 | 2023-11-10 | 摩尔线程智能科技(北京)有限责任公司 | Network training method and device for posture estimation |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102074034B (en) * | 2011-01-06 | 2013-11-06 | 西安电子科技大学 | Multi-model human motion tracking method |
| DE102013203667B4 (en) * | 2013-03-04 | 2024-02-22 | Adidas Ag | Cabin for trying out one or more items of clothing |
| CN106097278B (en) * | 2016-06-24 | 2021-11-30 | 北京工业大学 | Sparse model, reconstruction method and dictionary training method of multi-dimensional signal |
| CN106650827A (en) * | 2016-12-30 | 2017-05-10 | 南京大学 | Human body posture estimation method and system based on structure guidance deep learning |
-
2017
- 2017-08-18 CN CN201710712291.0A patent/CN107480720B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN107480720A (en) | 2017-12-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107480720B (en) | Human body posture model training method and device | |
| CN111104962B (en) | Semantic segmentation method and device for image, electronic equipment and readable storage medium | |
| CN108764292B (en) | Deep learning image target mapping and positioning method based on weak supervision information | |
| CN105095862B (en) | A kind of human motion recognition method based on depth convolution condition random field | |
| CN109360028B (en) | Method and device for pushing information | |
| CN108765506B (en) | Layer-by-layer network binarization-based compression method | |
| CN111476708B (en) | Model generation method, model acquisition method, device, equipment and storage medium | |
| KR20190055447A (en) | Apparatus and method for generating and using neural network model applying accelerated computation | |
| CN110717953B (en) | Coloring method and system for black-and-white pictures based on CNN-LSTM (computer-aided three-dimensional network-link) combination model | |
| CN112418292A (en) | Image quality evaluation method and device, computer equipment and storage medium | |
| CN111080655A (en) | Image segmentation and model training method, device, medium and electronic equipment | |
| CN106649782B (en) | Picture retrieval method and system | |
| CN117454495B (en) | CAD vector model generation method and device based on building sketch outline sequence | |
| CN107239532B (en) | Data mining method and device | |
| CN112883227A (en) | Video abstract generation method and device based on multi-scale time sequence characteristics | |
| JP2016029568A (en) | Linear identification device, large-sized general object recognition device, electronic computer, mobile terminal, data processor, and image recognition system | |
| CN116740362A (en) | Attention-based lightweight asymmetric scene semantic segmentation method and system | |
| US9471871B2 (en) | Method of generating features optimal to a dataset and classifier | |
| CN112257840B (en) | Neural network processing method and related equipment | |
| CN111325212A (en) | Model training method and device, electronic equipment and computer readable storage medium | |
| Dewan et al. | Image Classification by Transfer Learning using Pre-Trained CNN Models | |
| CN117933345B (en) | Training method of medical image segmentation model | |
| CN114764619B (en) | Convolution operation method and device based on quantum circuit | |
| CN106250873A (en) | Extracting of Moving Object and device | |
| CN115082430B (en) | Image analysis method and device and electronic equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |