A pose similarity recognition method, device and computer program product

Abstract

This disclosure provides a pose similarity recognition method, apparatus, and computer program product, relating to the field of image recognition technology. The pose similarity recognition method includes: acquiring the positional information of each target keypoint in two poses to be compared; based on the positional information, acquiring two pose features corresponding to the poses to be compared using a pre-trained pose feature extraction model; the pose feature extraction model is used to map the positional information to a feature vector space, such that pose features of similar poses are close to each other in the feature vector space, and pose features of dissimilar poses are far apart in the feature vector space; and based on the two pose features, determining the similarity between the two poses to be compared.

Description

Technical Field

[0001] This specification relates to the field of image recognition technology, and in particular to a pose similarity recognition method, device, and computer program product. Background Technology

[0002] With the rise of applications such as smart fitness, motion instruction, and motion-sensing interaction, human posture recognition and comparison have become fundamental for tasks such as motion scoring, posture correction, and motion retrieval. For recognizing the similarity of human postures, the Euclidean distance can be determined based on the coordinates of corresponding joints in two postures to determine whether the two postures are similar. However, due to differences in height or body shape among individuals, the coordinates of corresponding joints may vary significantly when the same posture is performed by different individuals, leading to a misjudgment of low similarity.

[0003] In view of this, some embodiments of this specification provide a pose similarity recognition method, apparatus, and computer program product, in order to improve the accuracy of pose similarity recognition between different individuals. Summary of the Invention

[0004] This specification provides one or more embodiments of a pose similarity recognition method, the method comprising: acquiring position information of each target key point in two poses to be compared; acquiring two pose features corresponding to the poses to be compared respectively through a pre-trained pose feature extraction model based on the position information; the pose feature extraction model being used to map the position information to a feature vector space, such that pose features of similar poses are close to each other in the feature vector space and pose features of dissimilar poses are far apart in the feature vector space; and determining the similarity between the two poses to be compared based on the two pose features.

[0005] According to one or more embodiments of this specification, a method for determining the similarity between two poses to be compared based on two pose features includes: determining a Euclidean distance based on the two pose features; determining a similarity score corresponding to the Euclidean distance based on a predetermined mapping relationship; the predetermined mapping relationship is used to indicate that the similarity score decreases monotonically as the Euclidean distance increases.

[0006] According to one or more embodiments of the present specification, the predetermined mapping relationship includes a first boundary value, a second boundary value, and a target threshold located between the first boundary value and the second boundary value; the first boundary value indicates that the two poses are completely similar, the second boundary value indicates that the two poses are completely dissimilar, and the target threshold indicates that the two poses are in a critical state of similarity and dissimilarity.

[0007] According to one or more embodiments of this specification, the Euclidean distance corresponding to the target threshold is determined by a validation sample set. Each validation sample in the validation sample set includes two first sample poses and a first sample label, where the first sample label indicates whether the two first sample poses are similar or dissimilar. The method further includes: obtaining first sample position information of each target keypoint in the two first sample poses of each validation sample; inputting the two first sample position information into a pre-trained pose feature extraction model to obtain two first sample pose features corresponding to the validation sample, and determining the first sample Euclidean distance between the two first sample pose features; using any first sample Euclidean distance as a reference distance, setting the validation samples corresponding to first sample Euclidean distances less than or equal to the reference distance as predicted similar samples, and combining the first sample labels of each validation sample to determine the true positive rate and false positive rate corresponding to the reference distance; the true positive rate represents the ratio of the number of correctly predicted samples in the predicted similar samples to the number of samples indicated by the first sample label ...

[0008] According to one or more embodiments of the method provided in this specification, the similarity score corresponding to the first boundary value is equal to 100, the similarity score corresponding to the second boundary value is equal to 0, and the similarity score corresponding to the target threshold is equal to 60.

[0009] According to one or more embodiments of this specification, the pose feature extraction model is trained using a training sample set. Each training sample in the training sample set includes two second sample poses and a second sample label, where the second sample label indicates whether the two second sample poses are similar or dissimilar. The pre-training process of the pose feature extraction model includes: obtaining the second sample position information of each target keypoint in the two second sample poses in the training sample; inputting the two second sample position information into the pose feature extraction model to obtain two second sample pose features; determining the second sample Euclidean distance between the two second sample pose features; determining the loss value of the loss function based on the second sample Euclidean distance and the second sample label in the training sample; adjusting the parameters of the pose feature extraction model using a backpropagation algorithm based on the loss value, and repeating the pre-training process until the loss value converges to obtain the pre-trained pose feature extraction model.

[0010] According to one or more embodiments of this specification, the training objective of the loss function includes: when two second sample poses are not similar, such that the loss value decreases to 0 as the Euclidean distance between the second samples increases; and when two second sample poses are similar, such that the loss value increases as the Euclidean distance between the second samples increases.

[0011] According to the method provided in one or more embodiments of this specification, the training sample set includes different types of poses, and the same type of pose is collected based on different heights and / or different body types.

[0012] According to one or more embodiments of this specification, the method obtains the position information of each target key point in two poses to be compared, including: obtaining the position coordinates of each target key point in two poses to be compared; obtaining two coordinate vectors corresponding to the poses to be compared based on the position coordinates, each coordinate vector containing a set of coordinate values ​​of each position coordinate of the corresponding pose to be compared, and representing the position information through the set of coordinate values.

[0013] According to one or more embodiments of this specification, the target key points include joints; the number of target key points in each pose to be compared is not less than 10; the number and position of target key points in two poses to be compared correspond.

[0014] This specification provides one or more embodiments of a pose similarity recognition device, the device comprising: a pose information acquisition module, configured to acquire position information of each target key point in two poses to be compared; a pose feature extraction module, configured to acquire two pose features corresponding to the poses to be compared, respectively, based on the position information and through a pre-trained pose feature extraction model; the pose feature extraction model is configured to map the position information to a high-dimensional feature space, such that pose features of similar poses are close to each other in the high-dimensional feature space and pose features of dissimilar poses are far apart in the high-dimensional feature space; and a similarity determination module, configured to determine the similarity between the two poses to be compared based on the two pose features.

[0015] One or more embodiments of this specification also provide a computer program product, including computer instructions or a computer program, which, when at least a portion of the computer instructions or computer program is executed by a processor, can implement the pose similarity recognition method described in some embodiments of this specification. Attached Figure Description

[0016] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. The same numbers in the drawings denote the same structures or steps.

[0017] Figure 1 This is a schematic diagram illustrating an application scenario of pose similarity recognition based on some embodiments of this specification.

[0018] Figure 2 This is an exemplary flowchart of a pose similarity recognition method according to some embodiments of this specification.

[0019] Figure 3This is an exemplary flowchart illustrating an approach to determining pose similarity according to some embodiments of this specification.

[0020] Figure 4 This is a schematic diagram illustrating a predetermined mapping relationship according to some embodiments of this specification.

[0021] Figure 5 This is an exemplary flowchart illustrating a method for determining a target threshold according to some embodiments of this specification.

[0022] Figure 6 This is an exemplary flowchart illustrating a training pose feature extraction model according to some embodiments of this specification.

[0023] Figure 7 This is a schematic diagram of a loss function for pose dissimilarities, as shown in some embodiments of this specification.

[0024] Figure 8 This is a schematic diagram of a loss function for pose similarity according to some embodiments of this specification.

[0025] Figure 9 This is an exemplary flowchart illustrating one method of obtaining location information according to some embodiments of this specification.

[0026] Figure 10 This is a schematic diagram of an attitude image according to some embodiments of this specification.

[0027] Figure 11 This is an exemplary block diagram of a pose similarity recognition device according to some embodiments of this specification. Detailed Implementation

[0028] To more clearly illustrate the technical solutions of the embodiments in this specification, the embodiments will be described in detail below with reference to the accompanying drawings. Obviously, the content described below are some examples or embodiments of this specification. For those skilled in the art, without creative effort, the technical solutions or means disclosed in this specification can be applied to other scenarios based on this technical content.

[0029] It should be understood that the terms "system," "device," "unit," and / or "module" used in this specification are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0030] Unless otherwise specified, the technical terms used to describe components, elements, etc. in this specification are not singular but may include plural. Generally speaking, terms such as "comprising" or "including" only indicate that explicitly identified steps, elements, or components are included, and these steps, elements, and components do not constitute an exclusive list, as the described method or apparatus may also include other steps or components.

[0031] This specification uses flowcharts to illustrate the operational steps performed by the apparatus or system of related embodiments. However, unless otherwise specified, the order in which these steps are described should not be construed as a limitation on the order of execution. Those skilled in the art can adjust the order of these steps based on the knowledge and information conveyed by the embodiments in this specification. Such adjustments include, but are not limited to, reversing the order of steps, merging multiple steps, and splitting a step.

[0032] With the rise of applications such as smart fitness, motion instruction, and motion-sensing interaction, human posture recognition and comparison have become fundamental for tasks such as motion scoring, posture correction, and motion retrieval. For recognizing the similarity of human postures, the Euclidean distance can be determined based on the coordinates of corresponding joints in two postures to determine whether the two postures are similar. However, due to differences in height or body shape among individuals, the coordinates of corresponding joints may vary significantly when the same posture is performed by different individuals, leading to a misjudgment of low similarity.

[0033] Therefore, some embodiments of this specification propose a posture similarity recognition method, apparatus, and computer program product to improve the accuracy of posture similarity recognition among different individuals. It is understood that in the specific implementations of this specification, the collection, use, or processing of data (such as postures to be compared, verification sample sets, training sample sets, etc.) is involved. When one or more embodiments of this specification are applied to specific products or technical implementations, permission or consent from the data subject is required. Furthermore, the collection, use, or processing of related data must strictly comply with the relevant laws, regulations, and standards of the data source country, implementation country, and other relevant countries and regions. De-identification technology ensures that the final data used is securely processed and de-identified data, protecting the rights and interests of the data subject and data security.

[0034] Figure 1 This is a schematic diagram illustrating an application scenario of pose similarity recognition according to some embodiments of this specification. In some embodiments, such as Figure 1As shown, application scenario 100 may include client 110, server 120, and network 130. Client 110 and server 120 can transmit data through network 130. Client 110 can provide an upload interface and upload entry for images such as verification samples, training samples, and poses to be compared, and send the user-uploaded verification samples, training samples, and poses to be compared to server 120. Both verification samples and training samples may include sample poses and sample labels. Each pair of sample poses corresponds to one sample label. Sample poses are input in pairs, and sample labels can be used to indicate whether two pairs of input sample poses are similar or dissimilar. Server 120 can train a pose feature extraction model for extracting pose features based on multiple training samples. It can also determine the target threshold of Euclidean distance corresponding to the critical state of similarity and dissimilarity between two poses based on multiple verification samples. Then, it can use the trained pose feature extraction model to obtain the corresponding pose features based on the poses to be compared, so as to determine the similarity between the two poses to be compared.

[0035] In some embodiments, client 110 may provide an upload interface and upload entry for images such as verification samples, training samples, and poses to be compared. In some embodiments, client 110 may be a terminal device including, but not limited to, desktop computers, smartphones, laptops, VR devices, tablets, smart TVs, and in-vehicle terminals. Client 110 may include a display screen and a processor. The display screen may be used to present a graphical user interface. For example, client 110 may present an upload interface and upload entry for images such as verification samples, training samples, and poses to be compared through the graphical user interface. In some embodiments, the display screen may be separate from the human-machine interface device. Users may operate on the graphical user interface through the human-machine interface device. The processor of client 110 may receive operation instructions generated by operations on the graphical user interface through the human-machine interface device. For example, the processor may generate a pose similarity recognition request based on the poses to be compared uploaded by the user, send the pose similarity recognition request to server 120, and receive the similarity comparison result fed back by server 120. In some embodiments, the display screen may be a touch screen, which may receive operation instructions input by the user based on the graphical user interface. In some embodiments, client 110 may include a memory for storing the received similarity comparison results.

[0036] Server 120 may be a high-performance computer device used to receive pose similarity recognition requests sent by client 110. The pose similarity recognition request may include two images corresponding to poses to be compared. Server 120 can use a trained pose feature extraction model to obtain corresponding pose features based on the poses to be compared, in order to determine the similarity between the two poses. In some embodiments, server 120 may include a local server or a cloud server. Depending on different service requirements, a local server or a cloud server corresponding to that region may be deployed in one or more regions. In some embodiments, server 120 may include a backend processing server. The backend processing server can train a pose feature extraction model for extracting pose features based on multiple training samples sent by client 110. It can also determine the target threshold corresponding to the critical state of similarity and dissimilarity between two poses based on multiple verification samples sent by client 110. Furthermore, it can use the trained pose feature extraction model to obtain corresponding pose features based on the poses to be compared sent by client 110, in order to determine the similarity between the two poses. In some embodiments, the server 120 may be a single computer device or a computing cluster composed of multiple computer devices, thereby providing more powerful computing power and more efficient data processing. For example, it may use a trained pose feature extraction model to quickly obtain the pose features corresponding to the poses to be compared and quickly determine the similarity between the two poses to be compared.

[0037] Network 130 can be any form of wired or wireless network, or any combination thereof. As an example, network 130 can be one or more combinations of wired networks, fiber optic networks, telecommunications networks, internal networks, the Internet, local area networks (LANs), wide area networks (WANs), wireless local area networks (WLANs), metropolitan area networks (MANs), public switched telephone networks (PSTNs), and Bluetooth networks. Network 130 can have multiple access points, and clients 110 and servers 120 can access network 130 through these access points.

[0038] It should be noted that, Figure 1 The illustrated application scenario diagram for pose similarity recognition is merely an example. The application scenarios for pose similarity recognition described in the embodiments of this specification are intended to more clearly illustrate the technical solutions of the embodiments of this specification and do not constitute a limitation on the technical solutions provided in the embodiments of this specification. For example, Figure 1The number of client 110 and server 120 is merely illustrative and not intended to limit the scope of patent protection of this application. Depending on the actual situation, there can be any number of client 110 and server 120. For example, client 110 and server 120 can be implemented as a single device, allowing for the uploading and recognition processing of images such as verification samples, training samples, and poses to be compared, as well as the training of the pose feature extraction model, all within the same device. As will be apparent to those skilled in the art, with the development of data processing technology and the emergence of new business scenarios, the technical solutions provided in the embodiments of this specification are also applicable to similar technical problems.

[0039] To improve the accuracy of recognizing the similarity of postures among different individuals, this specification provides a posture similarity recognition method. Figure 2 This is an exemplary flowchart of a pose similarity recognition method according to some embodiments of this specification. Figure 2 The process 200 shown can be executed by a processing device, for example, by... Figure 1 The server 120 shown executes the process. In some embodiments, process 200 can be implemented by a pose similarity recognition device 1100 deployed on a processing device. The pose similarity recognition device 1100 may include a pose information acquisition module 1110, a pose feature extraction module 1120, and a similarity determination module 1130. In some embodiments, such as Figure 2 As shown, process 200 may include the following steps.

[0040] Step 210: Obtain the position information of each target key point in the two poses to be compared. In some embodiments, step 210 can be implemented by the pose information acquisition module 1110.

[0041] In some embodiments, the poses to be compared can be two image data containing poses, such as human images captured by a camera.

[0042] In some embodiments, target keypoints may be spatial location points reflecting human posture, used to abstract and simplify human posture. In some embodiments, target keypoints may include joint-type keypoints, such as shoulder joints, elbow joints, wrist joints, hip joints, knee joints, ankle joints, etc. In some embodiments, target keypoints may include extremity-type keypoints, such as fingertips, top of the head, toes, etc. In some embodiments, target keypoints may include feature-type keypoints, such as the tip of the nose, palm, etc.

[0043] In some embodiments, location information may include the coordinates of target key points (e.g., two-dimensional or three-dimensional coordinates), or it may be represented by relative positions (e.g., using a fixed position as a reference point to determine the relative position information of each target key point relative to the reference point). In some embodiments, the coordinates of target key points may be extracted by a pose estimation model. For example, the pose estimation model may include the OpenPose (Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields) model, the HRNet (High-Resolution Network) model, etc.

[0044] In some embodiments, the positions of target keypoints in two poses to be compared are typically corresponding. In some embodiments, multiple target keypoints in each pose to be compared can be arranged according to the order of human body structure, and the order of the target keypoints at corresponding positions in the two poses to be compared is corresponding. The set of position information of each target keypoint in each pose to be compared can constitute a vectorized representation corresponding to the pose to be compared, serving as the basic data for subsequent pose similarity recognition.

[0045] In some embodiments, taking two-dimensional coordinates as an example for location information, if a pose has N target keypoints, the pose estimation model can extract N two-dimensional coordinates based on the pose image. Then, a pose can be represented by a 2N-dimensional vector. Assuming N=16, the pose can be represented as a 32-dimensional vector. .

[0046] Step 220: Based on the position information, two pose features corresponding to the poses to be compared are obtained through a pre-trained pose feature extraction model. The pose feature extraction model is used to map the position information to the feature vector space, and to make the pose features of similar poses close to each other in the feature vector space, and the pose features of dissimilar poses far apart in the feature vector space. In some embodiments, step 220 can be implemented by the pose feature extraction module 1120.

[0047] In some embodiments, the vector formed by the set of location information of each target key point can be used as the input of the attitude feature extraction model. The set of location information of two attitudes to be compared is input into the attitude feature extraction model respectively, and the attitude feature extraction model can output two attitude features corresponding to the two attitudes to be compared respectively.

[0048] In some embodiments, since the pose feature extraction model can map position information to the feature vector space, and make the pose features of similar poses close to each other in the feature vector space and the pose features of dissimilar poses far apart in the feature vector space, the pose features output by the pose feature extraction model can refer to the vectorized representation of the essential attributes of the pose extracted from the position information of the target key points. The essential attributes of the pose can refer to stable features that are directly related to the semantics of the action and are not affected by the body shape and shooting conditions.

[0049] In some embodiments, the pose features can be a high-dimensional semantic representation automatically extracted by the pose feature extraction model based on a set vector of location information of target key points, reflecting the semantic similarity between two poses to be compared.

[0050] In some embodiments, the pose feature extraction model may employ a convolutional neural network model.

[0051] Step 230: Determine the similarity between the two poses to be compared based on the two pose features. In some embodiments, step 230 can be implemented by the similarity determination module 1130.

[0052] In some embodiments, similarity can be used to measure the semantic proximity of two poses to be compared. In some embodiments, similarity can be presented as a percentage score, with a higher score indicating greater similarity between the two poses. Since pose features can reflect the essential attributes of a pose and reduce the impact of significant differences in the positional information of target key points caused by differences in height or body shape among different individuals, similarity determined based on pose features can more accurately reflect the proximity of two poses to be compared.

[0053] Figure 3 This is an exemplary flowchart illustrating an approach to determining pose similarity according to some embodiments of this specification. Figure 3 The process 300 shown can be executed by a processing device, for example, by... Figure 1 The server 120 shown executes this. In some embodiments, process 300 may be a further description of step 230. In some embodiments, process 300 may be implemented by a similarity determination module 1130 in a pose similarity recognition device 1100 deployed on a processing device. Figure 3 As shown, in some embodiments, process 300 may include the following steps.

[0054] Step 310: Determine the Euclidean distance based on the two pose features.

[0055] In some embodiments, Euclidean distance refers to the straight-line distance between two points in Euclidean space, which can be used to measure the proximity between two pose features. In some embodiments, the smaller the Euclidean distance, the closer the two pose features are in the feature vector space, i.e., the more semantically similar they are; the larger the Euclidean distance, the farther apart the two pose features are in the feature vector space, i.e., the greater the semantic difference and the less similar they are. Since Euclidean distance is determined based on posture features mapped by a posture feature extraction model, and since the posture feature extraction model is trained on human posture samples containing different heights, body types, and positions, the model parameters are continuously adjusted based on the loss function during training. This allows posture features of similar postures to move closer together in the feature vector space, while posture features of dissimilar postures to move further apart. This enables the model to filter out information irrelevant to posture semantics (such as the absolute coordinates of key points, limb length, etc.) and retain information that can characterize posture semantics, such as the relative positions between target key points determined based on the coordinates of the target key points and the relative angles formed by the lines connecting the target key points. Therefore, the distance of posture features in the feature vector space has reduced the interference of non-posture factors such as the height, body type, or image capture of different individuals through the posture feature extraction model. Thus, the determined Euclidean distance can more accurately reflect the semantic similarity of the posture itself.

[0056] Step 320: Based on the predetermined mapping relationship, determine the similarity score corresponding to the Euclidean distance; the predetermined mapping relationship is used to indicate that the similarity score decreases monotonically as the Euclidean distance increases.

[0057] In some embodiments, the predetermined mapping relationship can refer to a pre-defined function that converts Euclidean distance into a similarity score, defining the mapping relationship between Euclidean distance and similarity score. In some embodiments, the predetermined mapping relationship is used to indicate that the similarity score monotonically decreases with Euclidean distance, that is, as the Euclidean distance increases, the similarity score decreases. This matches the situation where the greater the distance between two pose features in the feature vector space, the lower the similarity. By converting Euclidean distance into a similarity score, the similarity between two poses to be compared can be reflected more intuitively.

[0058] In some embodiments, the predetermined mapping relationship includes a first boundary value, a second boundary value, and a target threshold located between the first boundary value and the second boundary value; the first boundary value indicates that the two poses are completely similar, the second boundary value indicates that the two poses are completely dissimilar, and the target threshold indicates that the two poses are in a critical state of similarity and dissimilarity.

[0059] In some embodiments, the similarity score corresponding to the first boundary value is equal to 100, the similarity score corresponding to the second boundary value is equal to 0, and the similarity score corresponding to the target threshold is equal to 60.

[0060] In some embodiments, the similarity score can be set to a percentage, with a similarity score of 100 for completely similar individuals, 0 for completely dissimilar individuals, and a passing score of 60 for individuals on the borderline between similarity and dissimilarity. In some embodiments, the similarity score corresponding to the target threshold can also be set to other values ​​as needed, such as 50. Figure 4 This is a schematic diagram illustrating a predetermined mapping relationship according to some embodiments of this specification. For example... Figure 4 As shown, for example, in the predetermined mapping relationship 400, the curve between the first boundary value and the target threshold has a steeper downward trend. As the Euclidean distance increases, the similarity score decreases rapidly, which can sensitively reflect the subtle differences between similar poses. The curve between the target threshold and the second boundary value has a gentler downward trend. As the Euclidean distance further increases, the similarity score decreases more slowly, which can avoid extreme changes in the similarity score (such as a slightly larger Euclidean distance but a rapid drop in the similarity score to zero), and preserve the distinguishability between dissimilar poses.

[0061] In some embodiments, the Euclidean distance corresponding to the target threshold is determined by a validation sample set. Each validation sample in the validation sample set includes two first sample poses and a first sample label. The first sample label is used to indicate whether the two first sample poses are similar or dissimilar. Figure 5 This is an exemplary flowchart illustrating a method for determining a target threshold according to some embodiments of this specification. Figure 5 The process 500 shown can be executed by a processing device, for example, by... Figure 1 The server 120 shown is executed. Figure 5 As shown, in some embodiments, process 500 may include the following steps.

[0062] Step 510: Obtain the first sample position information of each target key point in the pose of the two first samples in each verification sample.

[0063] In some embodiments, the verification samples consist of first sample pose pairs, each first sample pose pair corresponding to a first sample label. In some embodiments, the first sample location information may include the coordinates of target keypoints (e.g., two-dimensional or three-dimensional coordinates), or it may be represented by relative positions (e.g., using a fixed position as a reference point to determine the relative position information of each target keypoint relative to the reference point). In some embodiments, taking two-dimensional coordinates as an example, the two-dimensional coordinates of each target keypoint in the first sample pose can be extracted using a pose estimation model to obtain the first sample location information.

[0064] Step 520: Input the position information of the two first samples into the pre-trained pose feature extraction model to obtain the pose features of the two first samples corresponding to the validation samples, and determine the first sample Euclidean distance between the two first sample pose features.

[0065] In some embodiments, each validation sample includes a pair of first sample poses consisting of two first sample poses, with each pair of first sample poses corresponding to a first sample label. The resulting two first sample pose features correspond to the two first sample poses in the validation sample, respectively. In some embodiments, taking position coordinates as an example, the vector corresponding to the first sample pose can be determined based on the coordinate set of each target keypoint in each first sample pose. The vectors corresponding to the two first sample poses in each validation sample are grouped together and input into a pre-trained pose feature extraction model. The pose feature extraction model outputs two high-dimensional vectors, which correspond to the first sample pose features of the two first sample poses, respectively.

[0066] In some embodiments, the first sample Euclidean distance is used to measure the proximity between two first sample pose features. In some embodiments, for each pair of first sample poses in the validation samples, the process of determining the first sample Euclidean distance in step 520 is executed to determine the first sample Euclidean distance corresponding to the first sample pose pair. That is, multiple first sample Euclidean distances corresponding to the number of validation samples can be obtained in the end.

[0067] Step 530: Using any first sample Euclidean distance as a reference distance, set the validation samples corresponding to the first sample Euclidean distances that are less than or equal to the reference distance as predicted similar samples, and combine the first sample labels of each validation sample to determine the true positive rate and false positive rate corresponding to the reference distance.

[0068] The True Positive Rate (TPR) represents the ratio of the number of correctly predicted (actually similar validation samples are predicted as similar samples) to the number of samples whose first label indicates similarity among the predicted similar samples. The False Positive Rate (FPR) represents the ratio of the number of incorrectly predicted (actually dissimilar validation samples are predicted as similar samples) to the number of samples whose first label indicates dissimilarity among the predicted similar samples.

[0069] In some embodiments, the true positive rate reflects the proportion of all truly similar validation samples that are correctly identified as similar by the model. A higher expected true positive rate is better to avoid missing similar poses. In some embodiments, the false positive rate reflects the proportion of all truly dissimilar validation samples that are incorrectly identified as similar by the model. A lower expected false positive rate is better to avoid misclassifying dissimilar poses as similar.

[0070] In some embodiments, based on the Euclidean distances of the first samples determined by each validation sample, the validation samples can be sorted in ascending or descending order of their Euclidean distances to quickly identify the predicted similar samples corresponding to the first sample Euclidean distances that are less than or equal to the reference distance. By associating and storing the first sample labels with their corresponding first sample Euclidean distances, the number of correctly predicted (predicting actually similar validation samples as similar samples) and the number of incorrectly predicted (predicting actually dissimilar validation samples as similar samples) samples in the predicted similar samples can be quickly determined. Then, the true positive rate is determined based on the ratio of the number of correctly predicted samples to the number of samples indicated by the first sample labels as similar (actually similar), and the false positive rate is determined based on the ratio of the number of incorrectly predicted samples to the number of samples indicated by the first sample labels as dissimilar (actually dissimilar), thereby improving the computational efficiency of the true positive rate and the false positive rate.

[0071] Step 540: The reference distance corresponding to the maximum difference between the true positive rate and the false positive rate is taken as the Euclidean distance corresponding to the target threshold.

[0072] In some embodiments, since a higher expected true positive rate is better and a lower expected false positive rate is better, the maximum difference between the true positive rate and the false positive rate indicates that the model can identify most truly similar poses and will not misclassify too many dissimilar poses as similar. Using the reference distance corresponding to the maximum difference between the true positive rate and the false positive rate as the Euclidean distance corresponding to the target threshold (representing the critical state between similarity and dissimilarity between two poses, with a similarity score of, for example, equal to 60) avoids extreme cases. For example, if the reference distance corresponding to the maximum true positive rate is set as the Euclidean distance corresponding to the target threshold, almost all validation samples would be judged as similar; conversely, if the reference distance corresponding to the maximum false positive rate is set as the Euclidean distance corresponding to the target threshold, almost all validation samples would be judged as dissimilar. Using the reference distance corresponding to the maximum difference between the true positive rate and the false positive rate as the Euclidean distance of the target threshold, the model can find a balance between not missing positive examples (positive examples represent actual similar samples, missing positive examples means that the predicted similar samples omit some actual similar samples) and not misclassifying negative examples (negative examples represent actual dissimilar samples, misclassifying negative examples means that the predicted similar samples include some actual dissimilar samples). This results in Euclidean distances less than the target threshold being judged as pose similar, and Euclidean distances greater than the target threshold being judged as dissimilar. Combined with the aforementioned predetermined mapping relationship, an intuitive similarity score can be obtained based on the Euclidean distance transformation.

[0073] In some embodiments, the pose feature extraction model is trained using a training sample set, where each training sample includes two second sample poses and a second sample label, the second sample label being used to indicate whether the two second sample poses are similar or dissimilar. Figure 6 This is an exemplary flowchart illustrating a training pose feature extraction model according to some embodiments of this specification. Figure 6 The pre-training process 600 of the pose feature extraction model shown can be executed by a processing device, for example, by... Figure 1 The server 120 shown is executed. Figure 6 As shown, in some embodiments, the pre-training process 600 may include the following steps.

[0074] Step 610: Obtain the second sample position information of each target key point in the two second sample poses of the training sample.

[0075] In some embodiments, the training sample set may include different types of postures. The type of posture can refer to the semantic category of human action posture, such as hands on hips, raising hands, kicking legs, standing, etc. The same type of posture is collected based on different heights and / or different body types. In some embodiments, in order for the trained posture feature extraction model to extract posture features of different postures, the training sample set used to train the posture feature extraction model may contain diverse posture types. Each type of posture is collected based on people of different heights and / or different body types. For people of different heights and / or different body types, when their postures are similar, such as both being in the hands-on-hips posture, the posture feature extraction model can identify them as similar postures.

[0076] In some embodiments, the training samples consist of second sample pose pairs, with each second sample pose pair corresponding to a second sample label. In some embodiments, the second sample location information may include the coordinates of target keypoints (e.g., two-dimensional or three-dimensional coordinates), or it may be represented by relative positions (e.g., using a fixed position as a reference point to determine the relative position information of each target keypoint relative to the reference point). In some embodiments, taking two-dimensional coordinates as an example, the two-dimensional coordinates of each target keypoint in the second sample pose can be extracted using a pose estimation model to obtain the second sample location information.

[0077] Step 620: Input the position information of the two second samples into the pose feature extraction model to obtain the pose features of the two second samples.

[0078] In some embodiments, taking the position coordinates of the second sample position information as an example, the vector corresponding to the second sample posture can be determined based on the set of coordinate values ​​of each target key point in each second sample posture. The vectors corresponding to the two second sample postures in each training sample are used as a group and input to the posture feature extraction model to be trained. The posture feature extraction model outputs two high-dimensional vectors, which are the second sample posture features corresponding to the second sample posture. For training samples with different heights and / or different body types, the more similar the two second sample postures are, the closer the two high-dimensional vectors (second sample posture features) output by the posture feature extraction model are in the feature vector space.

[0079] Step 630: Determine the Euclidean distance between the pose features of the two second samples.

[0080] In some embodiments, the second-sample Euclidean distance is used to measure the proximity between two second-sample pose features. For training samples of different heights and / or body types, the more similar the poses of two second-sample samples, the closer the two high-dimensional vectors (second-sample pose features) output by the pose feature extraction model are in the feature vector space, and the smaller the second-sample Euclidean distance determined based on the two second-sample pose features. A smaller second-sample Euclidean distance means that the two second-sample poses are more similar.

[0081] Step 640: Determine the loss value of the loss function based on the Euclidean distance of the second sample and the label of the second sample in the training samples.

[0082] In some embodiments, the second sample label is a pre-labeled supervision signal used to indicate whether two second sample poses in the training samples are similar or dissimilar. In some embodiments, the second sample label can be 1, indicating that the two second sample poses belong to the same category, that is, they are semantically similar. For example, the semantics corresponding to the "hands-on-hips" category can be that the wrist joint is close to the hip joint. If the semantic information obtained based on the two second sample poses is that the wrist joint is close to the hip joint, then the two second sample poses can be considered to belong to the "hands-on-hips" category. The second sample label can also be 0, indicating that the two second sample poses belong to different categories, that is, they are semantically dissimilar.

[0083] In some embodiments, the loss function can be used to measure the difference between the model's prediction and the true label. During model training, the loss value of the loss function can guide the optimization direction of the model parameters. In some embodiments, the loss value refers to the numerical result calculated by substituting the specific second sample label and the determined second sample Euclidean distance into the loss function, reflecting the degree of matching between the pose feature extraction model's prediction of the training samples and the true label under the current parameters. In some embodiments, the smaller the loss value, the closer the model prediction is to the expectation (a smaller Euclidean distance for second samples based on similar training samples, and a larger Euclidean distance for second samples based on dissimilar training samples); the larger the loss value, the farther the model prediction deviates from the expectation.

[0084] In some embodiments, for example, taking the second sample label corresponding to similar training samples as 1 and the second sample label corresponding to dissimilar training samples as 0, assuming the parameter corresponding to the second sample label is label, where label represents the second sample label and its value is equal to 1 (indicating similarity) or 0 (indicating dissimilarity), and the parameter corresponding to the Euclidean distance of the second samples is dis, then the loss function can be constructed as follows: For similar samples (label=1), the loss value (Loss) increases as the Euclidean distance (dis) of the second sample increases; for dissimilar samples (label=0), the loss value (Loss) decreases to 0 as the Euclidean distance (dis) of the second sample increases.

[0085] Step 650: Based on the loss value, adjust the parameters of the pose feature extraction model through the backpropagation algorithm, and repeat the pre-training process until the loss value converges to obtain the pre-trained pose feature extraction model.

[0086] In some embodiments, the backpropagation algorithm is used to train a model, which calculates the gradient (partial derivative) of the loss function with respect to the model parameters and updates the model parameters based on the gradient. In some embodiments, the model parameters are learnable variables within the model that determine how the model maps inputs to outputs. These parameters are continuously adjusted during the model's training process, resulting in a gradual improvement in the model's performance.

[0087] In some embodiments, loss convergence refers to the state in which the loss value gradually decreases and eventually stabilizes as the number of iterations increases during the pre-training process. Loss convergence can serve as a sign that the model training is complete.

[0088] In some embodiments, the training sample set contains a large number of training samples. By continuously inputting different training samples into the pose feature extraction model and repeating the pre-training process, the performance of the pose feature extraction model can be continuously improved until the loss value converges. It can be considered that the pose feature extraction model corresponding to the current parameters has been trained. Regardless of whether the training samples are collected from human poses of different heights and / or different body types, the trained pose feature extraction model has the ability to map the position information of each target key point in the pose to the feature vector space, and to make the pose features of similar poses close to each other in the feature vector space and the pose features of dissimilar poses far away from each other in the feature vector space.

[0089] In some embodiments, the training objective of the loss function may include: when two second sample poses are dissimilar, causing the loss value to decrease to 0 as the Euclidean distance between the second samples increases. In some embodiments, taking a second sample label of 0 corresponding to dissimilar training samples as an example, substituting the second sample label=0 into the aforementioned constructed loss function simplifies the loss function to: . Figure 7 This is a schematic diagram illustrating a loss function for pose dissimilarity according to some embodiments of this specification, such as... Figure 7 As shown, in the loss function 700 under dissimilarity, when the two poses are dissimilarity, it is desirable for the Euclidean distance dis of the second sample to be as high as possible. Therefore, the expected loss value decreases as the Euclidean distance dis of the second sample increases until it becomes 0.

[0090] In some embodiments, the training objective of the loss function may include: when two second sample poses are similar, the loss value increases with the increase of the Euclidean distance between the second samples. In some embodiments, taking the second sample label corresponding to similar training samples as 1 as an example, corresponding to the aforementioned constructed loss function, substituting the second sample label label=1 into the loss function, the loss function can be simplified to: . Figure 8 This is a schematic diagram of a loss function for pose similarity according to some embodiments of this specification, such as... Figure 8 As shown, in the loss function 800 under similar conditions, when two poses are similar, it is desirable to have a lower Euclidean distance dis for the second sample. Therefore, the expected loss value increases as the Euclidean distance dis for the second sample increases.

[0091] Figure 9 This is an exemplary flowchart illustrating one method of obtaining location information according to some embodiments of this specification. Figure 9 The process 900 shown can be executed by a processing device, for example, by... Figure 1The server 120 shown executes this. In some embodiments, process 900 may be a further description of step 210. In some embodiments, process 900 may be implemented by the pose information acquisition module 1110 in the pose similarity recognition device 1100 deployed on the processing device. Figure 9 As shown, in some embodiments, process 900 may include the following steps.

[0092] Step 910: Obtain the position coordinates of each target key point in the two poses to be compared.

[0093] In some embodiments, the location coordinates may include two-dimensional coordinates or three-dimensional coordinates.

[0094] In some embodiments, target keypoints may include joints. In some embodiments, based on each target keypoint, two adjacent target keypoints that have a connection relationship (e.g., knee joint and ankle joint) can be connected by lines to construct a simplified pose image. Figure 10 This is a schematic diagram of an attitude image according to some embodiments of this specification, such as... Figure 10 As shown, the pose image set 1000 contains simplified pose images of different types. The type of pose can refer to the semantic category of human action posture, such as hands on hips (wrist joint close to hip joint), hands raised (wrist joint higher than elbow joint), etc.

[0095] In some embodiments, the number of target keypoints in each pose to be compared is no less than 10. Too few target keypoints may make it impossible to distinguish between different poses, while too many target keypoints require more computational resources. Therefore, an appropriate number of target keypoints can be selected based on actual accuracy requirements and computational performance.

[0096] In some embodiments, the number and position of target key points in two poses to be compared correspond, so that the position information subsequently obtained based on the target key points corresponds, thereby making them comparable.

[0097] Step 920: Based on the position coordinates, two coordinate vectors corresponding to the attitudes to be compared are obtained respectively. Each coordinate vector contains a set of coordinate values ​​of each position coordinate of the corresponding attitude to be compared, and the position information is represented by the set of coordinate values.

[0098] In some embodiments, assuming that the number of target key points in each pose to be compared is N, taking two-dimensional coordinates as an example, the coordinate vector can be a 2N-dimensional vector, and taking three-dimensional coordinates as an example, the coordinate vector can be a 3N-dimensional vector.

[0099] This specification also provides a posture similarity recognition device. Figure 11This is an exemplary block diagram of a pose similarity recognition device according to some embodiments of this specification. In some embodiments, the pose similarity recognition device 1100 may be deployed in a processing device. In some embodiments, the pose similarity recognition device 1100 may be deployed on a server 120. Figure 11 As shown, in some embodiments, the posture similarity recognition device 1100 may include a posture information acquisition module 1110, a posture feature extraction module 1120, and a similarity determination module 1130.

[0100] In some embodiments, the attitude information acquisition module 1110 can be used to acquire the position information of each target key point in two attitudes to be compared.

[0101] In some embodiments, the attitude information acquisition module 1110 can also be used to acquire the position coordinates of each target key point in the two attitudes to be compared; obtain two coordinate vectors corresponding to the attitudes to be compared based on the position coordinates, each coordinate vector containing a set of coordinate values ​​of each position coordinate of the corresponding attitude to be compared, and represent the position information through the set of coordinate values.

[0102] In some embodiments, the pose feature extraction module 1120 can be used to obtain two pose features corresponding to the poses to be compared respectively through a pre-trained pose feature extraction model based on position information; the pose feature extraction model is used to map the position information to a high-dimensional feature space, and make the pose features of similar poses close to each other in the high-dimensional feature space and the pose features of dissimilar poses far apart in the high-dimensional feature space.

[0103] In some embodiments, the similarity determination module 1130 can be used to determine the similarity between two poses to be compared based on two pose features.

[0104] In some embodiments, the similarity determination module 1130 can also be used to: determine the Euclidean distance based on two pose features; determine the similarity score corresponding to the Euclidean distance based on a predetermined mapping relationship; the predetermined mapping relationship is used to indicate that the similarity score decreases monotonically with the Euclidean distance.

[0105] For more information on each module, please refer to [link / reference]. Figures 2 to 10 The relevant explanations will not be repeated here. It should be understood that... Figure 11The modules shown can be implemented in various ways. For example, in some embodiments, the system and its modules can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated hardware. Those skilled in the art will understand that the methods and systems described above can be implemented using computer-executable instructions and / or included in the control code of a processor, such as on a media such as a disk, CD, or DVD-ROM, or in the memory of a programmable device. The systems and modules of this specification can be implemented not only with hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., but also with software, for example, executed by various types of processors, or with a combination of the aforementioned hardware circuits and software (e.g., firmware).

[0106] It should be noted that the above description of the system and its modules is for convenience only and should not be construed as limiting this specification to the embodiments described. It is understood that those skilled in the art, after understanding the principles of this system, may arbitrarily combine the various modules without departing from these principles to form subsystems connected to other modules. Alternatively, some modules may be split to obtain more modules or multiple units under a single module. Such modifications are all within the scope of this specification.

[0107] Some embodiments of this specification also provide a computer program product, including computer instructions or a computer program, which, when at least a portion of the computer instructions or computer program is executed by a processor, can implement this specification. Figures 2 to 10 The illustrated pose similarity recognition method. In some embodiments, the computer program product may involve only a computer program, which may be carried on a storage medium or processing device. In other embodiments, the computer program product may also be a storage medium or processing device containing a computer program. The processing device may include one or more processors, and the storage medium.

[0108] In some embodiments, the processor may be a combination of one or more of the following processors: central processing unit (CPU), application-specific integrated circuit (ASIC), application-specific instruction set processor (ASIP), graphics processing unit (GPU), physical processing unit (PPU), digital signal processor (DSP), field-programmable gate array (FPGA), programmable logic device (PLD), programmable logic controller (PLC), reduced instruction set computer (RISC), and microprocessor.

[0109] In some embodiments, the storage medium may include one or more combinations of the following: mass storage, removable storage, volatile read-write memory, and read-only memory (ROM). Exemplary mass storage may include disks, optical disks, solid-state drives, etc. Exemplary removable storage may include flash drives, floppy disks, optical disks, memory cards, compressed hard disks, magnetic tapes, etc. Exemplary volatile read-write memory may include random access memory (RAM). Exemplary RAM may include dynamic random access memory (DRAM), dual data rate synchronous dynamic random access memory (DDRSDRAM), static random access memory (SRAM), silicon controlled retrieval memory (T-RAM), and zero-capacitance memory (Z-RAM), etc. Exemplary read-only memory may include masked read-only memory (MROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), compressed hard disk read-only memory (CD-ROM), and digital multifunction hard disk read-only memory, etc.

[0110] The beneficial effects that the embodiments of this specification may bring include, but are not limited to: (1) by obtaining the position information of each target key point in two poses to be compared, the position information is mapped to the feature vector space by a pre-trained pose feature extraction model, so that the pose features of similar poses in the feature vector space are close to each other and the pose features of dissimilar poses in the feature vector space are far apart. The pose feature extraction model reduces the interference of different individuals' height and / or body shape differences on the pose. Then, the similarity of the poses to be compared is determined based on the mapped pose features, which can improve the accuracy of similarity judgment; (2) by determining the similarity score corresponding to the Euclidean distance based on a predetermined mapping relationship, the two poses can be compared. The similarity is presented more intuitively in the form of similarity scores; (3) By collecting the same type of posture based on different heights and / or different body types, and constructing a training sample set based on different types of postures, for each posture action (such as hands on hips, raising hands, etc.), collect multiple images of people of different heights and / or different body types performing the posture action, take two images of the same type of posture (such as two images of hands on hips) as a similar training sample, and take two images of different types of posture (such as one image of hands on hips and one image of raising hands) as a dissimilar training sample, so that the posture feature extraction model can be trained based on diverse training samples, thereby improving the model's ability to recognize diverse postures. It should be noted that different embodiments may produce different beneficial effects. In different embodiments, the beneficial effects that may be produced can be any one or a combination of the above, or any other possible beneficial effects.

[0111] The basic concepts have been described above. It is obvious that the detailed disclosure above is merely illustrative and does not constitute a limitation of this specification. Although not explicitly stated herein, various modifications, improvements, and corrections may be made to this specification by those skilled in the art. Such modifications, improvements, and corrections are taught in this specification and therefore remain within the spirit and scope of the exemplary embodiments described herein.

Claims

1. A pose similarity recognition method, characterized in that, The method includes: Obtain the position information of each target key point in the two poses to be compared; Based on the location information, two pose features corresponding to the poses to be compared are obtained by a pre-trained pose feature extraction model. The pose feature extraction model is used to map the location information to a feature vector space, so that the pose features of similar poses in the feature vector space are close to each other and the pose features of dissimilar poses in the feature vector space are far apart. Based on the two posture features, the similarity between the two postures to be compared is determined.

2. The method according to claim 1, characterized in that, Determining the similarity between two poses to be compared based on the two pose features includes: The Euclidean distance is determined based on the two aforementioned attitude features; Based on a predetermined mapping relationship, a similarity score corresponding to the Euclidean distance is determined; the predetermined mapping relationship is used to indicate that the similarity score decreases monotonically as the Euclidean distance increases.

3. The method according to claim 2, characterized in that, The predetermined mapping relationship includes a first boundary value, a second boundary value, and a target threshold located between the first boundary value and the second boundary value; the first boundary value indicates that the two poses are completely similar, the second boundary value indicates that the two poses are completely dissimilar, and the target threshold indicates that the two poses are in a critical state of similarity and dissimilarity.

4. The method according to claim 3, characterized in that, The Euclidean distance corresponding to the target threshold is determined by a validation sample set. Each validation sample in the validation sample set includes two first sample poses and a first sample label. The first sample label is used to indicate whether the two first sample poses are similar or dissimilar. The method further includes: Obtain the first sample position information of each target key point in the pose of two first samples in each of the verification samples; The position information of the two first samples is respectively input into the pre-trained pose feature extraction model to obtain two first sample pose features corresponding to the verification sample, and the first sample Euclidean distance between the two first sample pose features is determined. Using any of the first sample's Euclidean distances as a reference distance, the validation samples corresponding to the first sample's Euclidean distances that are less than or equal to the reference distance are set as predicted similar samples. Combining the first sample labels of each validation sample, the true positive rate and false positive rate corresponding to the reference distance are determined. The true positive rate represents the ratio of the number of correctly predicted samples to the number of samples indicated by the first sample label as similar, and the false positive rate represents the ratio of the number of incorrectly predicted samples to the number of samples indicated by the first sample label as dissimilar. The reference distance corresponding to the maximum difference between the true positive rate and the false positive rate is taken as the Euclidean distance corresponding to the target threshold.

5. The method according to claim 3, characterized in that, The similarity score corresponding to the first boundary value is 100, the similarity score corresponding to the second boundary value is 0, and the similarity score corresponding to the target threshold is 60.

6. The method according to any one of claims 1 to 5, characterized in that, The pose feature extraction model is trained using a training sample set. Each training sample in the training sample set includes two second sample poses and a second sample label. The second sample label is used to indicate whether the two second sample poses are similar or dissimilar. The pre-training process of the pose feature extraction model includes: Obtain the second sample position information of each target key point in two second sample poses in the training sample; The position information of the two second samples is respectively input into the pose feature extraction model to obtain the pose features of the two second samples; Determine the Euclidean distance between two second sample pose features; The loss value of the loss function is determined based on the Euclidean distance of the second sample and the label of the second sample in the training samples; Based on the loss value, the parameters of the pose feature extraction model are adjusted using the backpropagation algorithm, and the pre-training process is repeated until the loss value converges to obtain the pre-trained pose feature extraction model.

7. The method according to claim 6, characterized in that, The training objectives of the loss function include: when two second sample poses are not similar, the loss value decreases to 0 as the Euclidean distance between the second samples increases; and when two second sample poses are similar, the loss value increases as the Euclidean distance between the second samples increases.

8. The method according to claim 6, characterized in that, The training sample set includes different types of poses, and the same type of pose is collected based on different heights and / or different body types.

9. The method according to any one of claims 1 to 8, characterized in that, The step of acquiring the position information of each target key point in the two poses to be compared includes: Obtain the position coordinates of each of the target key points in the two postures to be compared; Based on the position coordinates, two coordinate vectors corresponding to the attitudes to be compared are obtained respectively. Each coordinate vector contains a set of coordinate values ​​of each position coordinate of the attitude to be compared, and the position information is represented by the set of coordinate values.

10. The method according to claim 9, characterized in that, The target key points include joints; the number of target key points in each of the postures to be compared is not less than 10; The number and position of the target key points in the two poses to be compared correspond.

11. A posture similarity recognition device, characterized in that, The device includes: The attitude information acquisition module is used to acquire the position information of each target key point in the two attitudes to be compared. The pose feature extraction module is used to obtain two pose features corresponding to the poses to be compared, respectively, through a pre-trained pose feature extraction model based on the position information. The pose feature extraction model is used to map the position information to a high-dimensional feature space, and to make the pose features of similar poses close to each other in the high-dimensional feature space and the pose features of dissimilar poses far apart from each other in the high-dimensional feature space. A similarity determination module is used to determine the similarity between two poses to be compared based on two pose features.

12. A computer program product, characterized in that, It includes computer instructions or computer programs, which, when executed by a processor in at least a portion, enable the pose similarity recognition method as described in any one of claims 1 to 10.

Images