Lightweight human key point detection method based on deep learning

Abstract

The application relates to a lightweight human key point detection method based on deep learning, which builds a detection network by taking MobileNetV3 as a main part, inputs a picture of W*H*C, carries out deep separable convolution and average pooling and h-swish activation function, finally outputs a feature map of 1*k, replaces an SE module in the MobileNetV3 with a PSA channel attention module, can better extract multi-scale spatial information, uses Adam to optimize training parameters, trains the network on a COCO dataset, then calculates a key point heat map and a partial affinity field of the dataset picture, estimates the calculated partial affinity field and the key point heat map and a key point label previously prepared on a test set, and calculates the accuracy. The application solves the problems that a large amount of memory is occupied during feature reasoning and the robustness is not strong under low-precision calculation, and the human key point position cannot be accurately predicted in a complex scene, and compared with existing lightweight human key point detection models, the application has certain competitiveness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a lightweight human key point detection method based on deep learning. BACKGROUND

[0002] Human key point detection is an important research direction in the field of computer vision, and has important practicality in human-computer interaction, monitoring security, sports competition. In recent years, with the development of deep learning, traditional geometric algorithms based on template matching have been gradually replaced by deep learning methods. The "bottom-up" detection method in deep learning is to predict human key points and then splice them. The traditional lightweight human key point detection method uses MobileNetV1 to replace VGG19 in Openpose as the backbone network, but MobileNetV1 not only occupies a large amount of memory during feature reasoning, but also has poor robustness under low precision calculation, making it difficult to accurately predict the position of human key points in complex scenes. SUMMARY

[0003] The present application aims to overcome the defects of the prior art and provide a lightweight human key point detection method based on deep learning, which solves the problems of large memory occupation, low robustness and difficulty in predicting human key points in complex scenes during traditional human pose estimation. While ensuring a small amount of calculation parameters, the spatial information of different scale feature maps is also used to enrich the feature space and enhance the prediction accuracy of human key points.

[0004] The technical solution to achieve the above-mentioned purpose is:

[0005] The present application provides a lightweight human key point detection method based on deep learning, comprising the following steps:

[0006] Provide pictures for training, build a detection network, input the pictures into the detection network to perform multi-scale feature extraction and fusion on the pictures, and add improved channel attention to output the features of the pictures;

[0007] Provide a data set, train and optimize the training parameters of the detection network using the data set, save a plurality of training models and corresponding training losses at a set training frequency interval, and select the training model with the smallest training loss as the detection model;

[0008] Provide a test set, input the test set into the detection model to obtain a corresponding human key point heat map, and then calculate the accuracy of the detection model;

[0009] When building the detection network, it further comprises:

[0010] MobileNetV3 is taken as the main part of the detection network, and the picture is input into the detection network, and the feature map is output through the depth separable convolution, the average pooling and the h-swish activation function;

[0011] The picture is input into the detection network, and a 3*3 convolution operation is performed on the picture, the output channel number of the picture is adjusted to 16, and standard normalization and h-swish nonlinear activation are performed to obtain a picture P1;

[0012] The picture P1 is subjected to a 3*3 depth separable convolution operation, and a PSA channel attention module is introduced, and standard normalization and ReLu nonlinear activation are performed to obtain a picture P2 with a channel number of 16;

[0013] The picture P2 is subjected to a 3*3 depth separable convolution operation, and then normalized and ReLu nonlinear activated to obtain a picture P3 with a channel number of 24;

[0014] The picture P3 is subjected to a 3*3 depth separable convolution operation, and then normalized and ReLu nonlinear activated to obtain a picture P4 with a channel number of 24;

[0015] The picture P4 is subjected to a 5*5 depth separable convolution operation, and a PSA channel attention module is introduced, and standard normalization and h-swish nonlinear activation are performed to obtain a picture P5 with a channel number of 40;

[0016] The picture P5 is subjected to a 5*5 depth separable convolution operation, and a PSA channel attention module is introduced, and standard normalization and h-swish nonlinear activation are performed to obtain a picture P6 with a channel number of 40;

[0017] The picture P6 is subjected to a 1*1 convolution operation, a PSA channel attention module is introduced, and standard normalization and h-swish nonlinear activation are performed to obtain a picture P7 with a channel number of 576;

[0018] The picture P7 is subjected to global average pooling, and the parameter quantity is reduced to obtain a picture P8 with a channel number of 576;

[0019] The picture P8 is input into a fully connected layer, and h-swish nonlinear activation is performed to obtain a picture P9 with a channel number of 1024;

[0020] The picture P9 is input into a fully connected layer, and finally a feature map with a channel number of k is obtained;

[0021] The improved attention also includes:

[0022] Input the feature map of W*H*C into the detection network, split the channels of the feature map by using the SPC module, then perform multi-scale feature extraction on the spatial information of each channel, and output the feature map F;

[0023] The feature map F is used to extract the channel attention vector under the feature map of different scales by using the SEWeight module;

[0024] The channel attention vectors of different scales are re-calibrated by using Softmax to obtain the corresponding attention weights;

[0025] The attention weights and the corresponding feature maps are point multiplied to obtain the weighted feature map of W*H*C.

[0026] The application proposes a lightweight human key point detection method based on deep learning, which builds a detection network by taking MobileNetV3 as the main part, inputs W*H*C picture, and finally outputs 1*k feature map through depth separable convolution, average pooling and h-swish activation function, proposes PSA channel attention module to replace SE module in MobileNetV3, which can better extract multi-scale spatial information, uses Pytorch framework to build neural network, Adam optimizes training parameters, trains the network on COCO dataset, inputs test dataset into the network, calculates the key point heat map and part affinity field of the dataset picture, then estimates the calculated part affinity field and key point heat map with the pre-prepared key point label on the test set, and calculates the accuracy, solves the problems of large memory occupation during feature reasoning and low robustness in low precision calculation, and accurately predicts the position of human key points in complex scenes, and has certain competitiveness compared with existing lightweight human key point detection model.

[0027] Further improvement of the lightweight human key point detection method based on deep learning in the application is that the depth separable convolution operation further includes:

[0028] Input the feature map with N channels into the detection network, perform depth convolution operation, split N channels one by one, then use a convolution kernel for each channel to obtain a feature map (FM1, FM2, FM3, …, FM n ) with 1 channel, and then sequentially splice the obtained n feature maps with 1 channel to obtain an output feature map with N channels;

[0029] Perform point-by-point convolution operation, and perform convolution operation on the obtained output feature map with a 1*1 convolution kernel size, perform feature fusion on the feature maps with N channels, and obtain the fused feature map FM.

[0030] The further improvement of the lightweight human key point detection method based on deep learning of the application is that when the SPC module is used to split the channels of the feature map, the SPC module further comprises:

[0031] The feature map P is input into the SPC module, the feature map P is split into S parts, and multi-scale convolution is performed on each split feature map to extract spatial information of each split feature map at different scales.

[0032] The further improvement of the lightweight human key point detection method based on deep learning of the application is that when the SPC module is used to split the channels of the feature map, the SPC module further comprises:

[0033] The feature map F is input into the SEWeight module, global average pooling is used to generate channel statistics, and then global spatial information is embedded into the channel descriptor.

[0034] The further improvement of the lightweight human key point detection method based on deep learning of the application is that when the SPC module is used to split the channels of the feature map, the SPC module further comprises:

[0035] The Adam optimizer is set, the learning rate is set to 4e-5, and the weight decay is set to 5e-4.

[0036] The further improvement of the lightweight human key point detection method based on deep learning of the application is that when the SPC module is used to split the channels of the feature map, the SPC module further comprises:

[0037] After every 100 epochs, the training loss is printed, and after every 5000 epochs, the validation loss is verified and the corresponding training model is saved.

[0038] The further improvement of the lightweight human key point detection method based on deep learning of the application is that when the SPC module is used to split the channels of the feature map, the SPC module further comprises:

[0039] The detection model is used to obtain the partial affinity field of the picture in the test set, the calculated heat map and partial affinity field are compared with the pre-marked key point label on the picture, and the accuracy of the detection model is calculated. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The framework diagram of the lightweight human key point detection method based on deep learning of the application.

[0041] Figure 2 The framework structure diagram of the detection network in the lightweight human key point detection method based on deep learning of the application.

[0042] Figure 3 The schematic diagram of the improved attention module in the lightweight human key point detection method based on deep learning of the application.

[0043] Figure 4 The SPC module schematic diagram in the lightweight human key point detection method based on deep learning of the application.

[0044] Figure 5 The SEWeight module schematic diagram in the lightweight human key point detection method based on deep learning of the application. DETAILED DESCRIPTION

[0045] The application will be further described below in conjunction with the drawings and specific embodiments.

[0046] The application provides a lightweight human key point detection method based on deep learning, which builds a detection network by taking MobileNetV3 as a main part, inputs a picture of W*H*C, and finally outputs a feature map of 1*k through depth separable convolution, average pooling and h-swish activation function, proposes a PSA channel attention module to replace the SE module in MobileNetV3, can better extract multi-scale spatial information, builds a neural network by using a Pytorch framework, optimizes training parameters by using Adam, trains the network on a COCO dataset, inputs a test dataset into the network, calculates a key point heat map and a partial affinity field of the dataset picture, then estimates the calculated partial affinity field and key point heat map and a key point label previously prepared on the test set, and calculates the accuracy, solves the problems that a large amount of memory is occupied during feature reasoning and the robustness is not strong under low-precision calculation, and it is difficult to accurately predict the human key point position in a complex scene, and has certain competitiveness compared with existing lightweight human key point detection models. The lightweight human key point detection method based on deep learning of the application will be described below in conjunction with the drawings.

[0047] Reference Figure 1 The framework diagram of the lightweight human key point detection method based on deep learning of the application. The lightweight human key point detection method based on deep learning of the application will be described below in conjunction with Figure 1 The lightweight human key point detection method based on deep learning of the application will be described below in conjunction with

[0048] As Figures 1-5 The application provides a lightweight human key point detection method based on deep learning, which includes the following steps:

[0049] Pictures for training are provided, a detection network is built, the pictures are input into the detection network, multi-scale feature extraction and fusion of the pictures are performed, and an improved channel attention is added, so that the features of the pictures are output;

[0050] The data set is provided, the detection network is trained and the training parameters are optimized by using the data set, a plurality of training models and corresponding training losses are saved at a set training interval, and the training model with the minimum training loss is taken as the detection model;

[0051] The test set is provided, the test set is input into the detection model to obtain a heat map of the corresponding human key point, and then the accuracy of the detection model is calculated.

[0052] As a preferred embodiment of the present application, when building the detection network, further comprising:

[0053] MobileNetV3 is taken as the main part of the detection network, the picture is input into the detection network, the feature map is output through deep separable convolution, average pooling and h-swish activation function.

[0054] Further improvement of the light-weight human key point detection method based on deep learning of the present application is that when building the detection network, further comprising:

[0055] The picture is input into the detection network, a 3*3 convolution operation is performed on the picture, the output channel number of the picture is adjusted to 16, and standard normalization and h-swish nonlinear activation are performed to obtain the picture P1;

[0056] The picture P1 is subjected to a 3*3 deep separable convolution operation, and a PSA channel attention module is introduced, standard normalization and ReLu nonlinear activation are performed to obtain the picture P2 with a channel number of 16;

[0057] The picture P2 is subjected to a 3*3 deep separable convolution operation, and then normalized and ReLu nonlinear activated to obtain the picture P3 with a channel number of 24;

[0058] The picture P3 is subjected to a 3*3 deep separable convolution operation, and then normalized and ReLu nonlinear activated to obtain the picture P4 with a channel number of 24;

[0059] The picture P4 is subjected to a 5*5 deep separable convolution operation, and a PSA channel attention module is introduced, standard normalization and h-swish nonlinear activation are performed to obtain the picture P5 with a channel number of 40;

[0060] The picture P5 is subjected to a 5*5 deep separable convolution operation, and a PSA channel attention module is introduced, standard normalization and h-swish nonlinear activation are performed to obtain the picture P6 with a channel number of 40;

[0061] The picture P6 is subjected to a 1*1 convolution operation, a PSA channel attention module is introduced, and standard normalization and h-swish nonlinear activation are performed, to obtain a picture P7 with a channel number of 576;

[0062] The picture P7 is subjected to global average pooling and parameter reduction, to obtain a picture P8 with a channel number of 576;

[0063] The picture P8 is input into a fully connected layer and subjected to h-swish nonlinear activation, to obtain a picture P9 with a channel number of 1024;

[0064] The picture P9 is input into a fully connected layer, and finally a feature map with a channel number of k is obtained.

[0065] Specifically, the depth separable convolution operation further includes:

[0066] A feature map with a channel number of N is input into a detection network, subjected to a depth convolution operation, and then N channels are sequentially split, and a convolution kernel is used for each channel to obtain a feature map (FM1, FM2, FM3, …, FMN) with a channel number of 1, and then the n feature maps with a channel number of 1 are sequentially spliced to obtain an output feature map with a channel number of N; n ).

[0067] Point-by-point convolution operation is performed on the obtained output feature map, and the output feature map is subjected to a 1*1 convolution operation, and the N-channel feature map is subjected to feature fusion to obtain a fused feature map FM.

[0068] Preferably, the feature fusion includes: inputting a feature map with a channel number of N, assuming that its input is IN, the i-th channel of IN is denoted as IN_i, the output of the first layer of depth convolution is denoted as DCO1, and the i-th channel of DCO1 is denoted as DCO_i; the output of the second layer of depth convolution is denoted as DCO2, and the i-th channel of DCO2 is denoted as DCO2_i; the output of the third layer of depth convolution is denoted as DCO3, and the i-th channel of DCO3 is denoted as DCO3_i, and so on; since the two feature maps before and after the depth convolution operation are one-to-one corresponding in the output channel, DCO1_i is only related to IN_i, DCO2_i is only related to DCO1_i, DCO3_i is only related to DCO2_i, and so on; the purpose of introducing feature fusion is to break this corresponding relationship, so that the output of a certain channel can be used as the input of another channel through 1*1 convolution.

[0069] Further, the improved attention further includes:

[0070] A feature map with a size of W*H*C is input into a detection network, the channels of the feature map are split by using an SPC module, then multi-scale feature extraction is performed on the spatial information of each channel, and a feature map F is output.

[0071] The feature map F is used to extract channel attention vectors under different scale feature maps by using the SEWeight module;

[0072] The channel attention vectors of different scales are re-calibrated by using Softmax to obtain corresponding attention weights;

[0073] The attention weights and the corresponding feature maps are point multiplied to obtain a weighted W*H*C feature map.

[0074] Further, when the SPC module is used to split the channels of the feature map, it further includes:

[0075] The feature map P is input into the SPC module, and the feature map P is split into S parts, represented by [P1, P2, …, P S ], and the channel number of each split feature map is C i =C / S, P1, P2, …, P S The feature map is multi-scale convolution to extract the spatial information of each split feature map under different scales.

[0076] The specific calculation method of multi-scale feature extraction is as follows:

[0077] (1) ;

[0078] Wherein,

[0079] (2) ;

[0080] The feature map after multi-scale fusion:

[0081] (3)

[0082] Further, when the SEWeight module is used to extract channel attention vectors under different scale feature maps, it further includes:

[0083] The feature map F is input into the SEWeight module, and the channel statistics are generated by global average pooling, and then the global spatial information is embedded into the channel descriptor.

[0084] Specifically, the global average pooling operator calculation formula is as follows:

[0085] (4) ;

[0086] The attention weight of the cth channel in the SEWeight module can be written as:

[0087] (5) ;

[0088] wherein the symbol represents a linear unit operation, W0 and W1 represent fully connected layers, and the symbol represents an excitation function;

[0089] By extracting the attention weight information from the multi-scale preprocessed feature map, different scale attention weight vectors are obtained, and the expression of the attention weight vector is as follows:

[0090] (6)

[0091] Specifically, the Softmax re-scaling feature construction step is:

[0092] The extracted multi-scale channel attention weight is sent into the Softmax to realize information interaction between the multi-scale channel attention, and the formula for re-weighting the channel attention is as follows:

[0093] (7)

[0094] Further, when training the detection network, the method further comprises:

[0095] The Adam optimizer is set, the learning rate is set to 4e-5, and the weight decay is set to 5e-4.

[0096] Preferably, when training the detection network, the method further comprises:

[0097] After every 100 epochs, the training loss is printed, and after every 5000 epochs, the validation loss is verified and the corresponding training model is saved.

[0098] Further, when inputting the test set into the detection model, the method further comprises:

[0099] The partial affinity field of the picture in the test set is obtained by using the detection model, the calculated heat map and partial affinity field are compared with the pre-marked key point label on the picture, and the accuracy of the detection model is calculated.

[0100] The specific embodiments of the present application are as follows:

[0101] A lightweight human key point detection network framework based on deep learning is constructed, as shown in Figure 1 The basic structure of bneck in the framework is as shown in Figure 2 The network realizes the effect of output picture features by performing multi-scale feature extraction and fusion on the input picture and adding improved channel attention;

[0102] The neural network is constructed, with MobileNetV3 as the main component. The input image is a W*H*C image, which is processed by depthwise separable convolution, average pooling, and h-swish activation function, and finally outputs a 1*k feature map.

[0103] Among them, the attention module was improved, combined with Figure 3 As shown, the original SE module in MobileNetV3 is improved into a PSA pyramid segmentation attention module. The PSA module includes three sub-modules: 1) SPC module; 2) SEWeight module; 3) Softmax function;

[0104] 1) SPC module

[0105] Combination Figure 4 As shown, the most important part of the PSA channel attention module is the SPC module. It takes the input feature map P and divides P into S parts, using [P1, P2, ..., P...]. S Let C be the number of channels in each segmented feature map. i =C / S, then for P1,P2,…,P S The feature maps are used to extract spatial information of feature maps at different scales using multi-scale convolution, resulting in [F1, F2, ..., F...]. S Then, feature fusion is performed on the obtained S feature maps, and finally feature map F is output;

[0106] 2) SEWeight module

[0107] Combination Figure 5 As shown, the SEWeight module consists of two parts: squeezing and excitation, which are used to encode global information and adaptively recalibrate the relationship between channels, respectively. Typically, channel statistics can be generated by using global average pooling, which is used to embed global spatial information into the channel descriptor. When calculating attention weights, the feature map passes through two fully connected layers, which can more effectively combine linear information between channels. This facilitates the interaction between high-channel and low-channel dimensional information. The introduction of the excitation function can assign weights to the channels after channel interaction, thereby extracting information more effectively.

[0108] Given a W*H*C feature map, first perform a 1*1 ordinary convolution operation to change the number of channels in the feature map. Then, feed the output feature map P1 into a depthwise separable convolution. First, split the N channels of feature map P1 one by one, and then apply a convolution kernel to each channel to obtain feature maps with 1 channel (FM1, FM2, FM3, ..., FM). n), then the N feature maps with 1 channel obtained are spliced in sequence to obtain an output feature map with N channels, then a point-by-point convolution operation is performed, the output feature map obtained is subjected to a convolution operation with a 1*1 convolution kernel size, and the feature maps with N channels are subjected to feature fusion to obtain a fused feature map FM, the FM is taken as an input of the PSA channel attention module, then spatial information of the FM is subjected to multi-scale feature extraction, and an attention weight is calculated, the attention weight and a corresponding feature map are subjected to a point multiplication operation to obtain a final feature map W*H*C, then the feature map W*H*C is taken as an input of a next depth separable convolution, the above operation is repeated, finally the output feature map W*H*C is sent into an average pooling layer and two fully connected layers to obtain a final feature map with k channels;

[0109] Training the built network: the neural network is built by using the Pytorch framework, the Adam optimization training parameter is used, the network is trained on the COCO dataset, the training loss is printed after every 100 epochs, the validation loss is verified and the training model is saved after every 5000 epochs, finally the model with the minimum loss is saved as the best model;

[0110] Testing the trained network: the best trained model is loaded, the read data set is loaded into the built neural network, the key point heat map and the partial affinity field of the data set picture are calculated, then the key point label on the test set is estimated, and the accuracy is calculated.

[0111] The above embodiments of the application are described in detail in combination with the drawings, and those skilled in the art can make various changes to the application according to the above description. Therefore, some details in the embodiments should not constitute a limitation on the application, and the scope of protection of the application will be defined by the appended claims.

Claims

1. A lightweight human keypoint detection method based on deep learning, characterized in that, Includes the following steps: Provide images for training, build a detection network, input the images into the detection network to perform multi-scale feature extraction and fusion on the images, and add improved channel attention to output the features of the images; Provide a dataset, use the dataset to train the detection network and optimize the training parameters, save several training models and corresponding training losses at set training intervals, and use the training model with the smallest training loss as the detection model. A test set is provided, and the test set is input into the detection model to obtain the corresponding heat map of human key points, and then the accuracy of the detection model is calculated. The construction of the detection network also includes: Using MobileNetV3 as the main part of the detection network, the image is input into the detection network, and a feature map is output through depthwise separable convolution, average pooling, and h-swish activation function. The image is input into the detection network, a 3*3 convolution operation is performed on the image, the number of output channels of the image is adjusted to 16, and standard normalization and h-swish non-linear activation are performed to obtain image P1; The image P1 is subjected to a 3*3 depthwise separable convolution operation, and a PSA channel attention module is introduced to perform standard normalization and ReLU non-linear activation to obtain an image P2 with 16 channels. The image P2 is subjected to a 3*3 depthwise separable convolution operation, followed by normalization and ReLU non-linear activation to obtain an image P3 with 24 channels. The image P3 is subjected to a 3*3 depthwise separable convolution operation, followed by normalization and ReLU non-linear activation to obtain an image P4 with 24 channels. The image P4 is subjected to a 5*5 depthwise separable convolution operation, and a PSA channel attention module is introduced to perform standard normalization and h-swish non-linear activation to obtain an image P5 with 40 channels. The image P5 is subjected to a 5*5 depthwise separable convolution operation, and a PSA channel attention module is introduced to perform standard normalization and h-swish non-linear activation to obtain an image P6 with 40 channels. The image P6 is subjected to a 1*1 convolution operation, and the PSA channel attention module is introduced and normalized and h-swish non-linear activation is performed to obtain the image P7 with 576 channels. The image P7 is subjected to global average pooling, and the number of parameters is reduced to obtain an image P8 with 576 channels. The image P8 is input into a fully connected layer and h-swish nonlinear activation is performed to obtain an image P9 with 1024 channels; The image P9 is input into a fully connected layer to obtain a feature map with k channels. Improving attention also includes: The feature map of W*H*C is input into the detection network, the SPC module is used to segment the channels of the feature map, and then multi-scale feature extraction is performed on the spatial information of each channel, and the feature map F is output. The feature map F is used to extract channel attention vectors at different scales using the SEWeight module; The channel attention vectors at different scales are recalibrated using Softmax to obtain the corresponding attention weights; Perform a dot product operation between the attention weights and the corresponding feature maps to obtain a weighted W*H*C feature map.

2. The lightweight human keypoint detection method based on deep learning as described in claim 1, characterized in that, The depthwise separable convolution operation further includes: The feature map with N channels is input into the detection network, and a depthwise convolution operation is performed to split the N channels one by one. Then, a convolution kernel is applied to each channel to obtain a feature map with 1 channel (FM1, FM2, FM3, ..., FM). n Then, the n feature maps with 1 channel are sequentially concatenated to obtain an output feature map with N channels; Perform pointwise convolution operation, convolve the resulting output feature map with a 1*1 kernel size, fuse the feature maps of N channels to obtain the fused feature map FM.

3. The lightweight human keypoint detection method based on deep learning as described in claim 1, characterized in that, When using the SPC module to segment the channels of the feature map, it also includes: The feature map P is input into the SPC module, which divides the feature map P into S parts. Multi-scale convolution is performed on each segmented feature map to extract the spatial information of each segmented feature map at different scales.

4. The lightweight human keypoint detection method based on deep learning as described in claim 3, characterized in that, When using the SEWeight module to extract channel attention vectors from feature maps at different scales, it also includes: The feature map F is input into the SEWeight module, and channel statistics are generated through global average pooling, thereby embedding global spatial information into the channel descriptor.

5. The lightweight human keypoint detection method based on deep learning as described in claim 1, characterized in that, Training the detection network also includes: Configure the Adam optimizer with a learning rate of 4e-5 and a weight decay of 5e-4.

6. The lightweight human keypoint detection method based on deep learning as described in claim 5, characterized in that, Training the detection network also includes: Print the training loss after every 100 epochs, and verify the loss and save the corresponding training model after every 5000 epochs.

7. The lightweight human keypoint detection method based on deep learning as described in claim 1, characterized in that, When inputting the test set into the detection model, the following is also included: The detection model is used to obtain the partial affinity field of the images in the test set. The calculated heat map and partial affinity field are compared with the key point labels pre-marked on the images, and the accuracy of the detection model is calculated.

Images