Object key point positioning method, device and equipment

Abstract

The invention discloses an object key point positioning method, device and equipment. Comprising the following steps: inputting to-be-identified image data into an encoder of an image generation model to obtain potential image features; initializing a key point condition vector, and inputting the potential image features and the key point condition vector into a U-Net module of the image generation model to obtain updated potential image features and key point condition vectors; inputting the updated potential image features and the key point condition vectors into a key point thermodynamic diagram extraction module based on condition guidance to obtain a key point thermodynamic diagram; and determining key point coordinates of the object based on the key point thermodynamic diagram. According to the method provided by the invention, key point positioning is executed by utilizing generation prior in an image generation model, the thermodynamic diagram is generated by adopting a key point thermograph extraction module based on condition guidance, and then the thermodynamic diagram is optimized through thermograph expected loss. Therefore, the generalization ability and the positioning precision of the model are improved.

Description

Technical Field

[0001] The present application relates to the field of computer vision technology, and in particular to a method, device and equipment for locating key points of an object. Background Art

[0002] The rapid development of deep learning and computer vision is driving visual perception technology from coarse-grained target detection and positioning to more fine-grained object structure analysis, where the object structure can be represented by a series of predetermined key points (such as elbows, knees, etc.). As the basic technology for fine-grained object structure analysis, key point positioning technology has attracted the attention of academia and industry at home and abroad. Accurate and universal object key point technology can be applied to human-computer interaction, motion analysis, artificial intelligence and other fields, and has important research significance and application value.

[0003] Object key point localization technology aims to detect the key point coordinates of an object from an input image. According to the different ways of modeling key points, existing technologies can be divided into regression-based methods and heat map-based methods. However, these deep learning-based methods are essentially learning nonlinear mappings from images to target spaces from training data sets. A single, limited annotated training data set cannot provide a general object structure prior. Existing object key point localization technology trains neural networks on a single, small number of annotated data sets. The resulting model can only be applied to a test data set that is consistent with the distribution of the training data set. It is difficult to cope with complex and changeable scenarios in actual applications, and the generalization ability of the model cannot be guaranteed. Collecting massive amounts of annotated data to train object key point localization models can alleviate this problem, but it requires a lot of manpower and material resources, resulting in a huge waste of resources. Summary of the invention

[0004] The embodiments of the present application provide a method, device and equipment for locating key points of an object, so as to at least solve the technical problems of poor generalization ability and high training cost of the positioning model in the related art.

[0005] According to one aspect of an embodiment of the present application, a method for locating key points of an object is provided, comprising: Input the image data to be recognized into the encoder of the image generation model to obtain the potential image features; Initializing a key point condition vector, inputting the potential image feature and the key point condition vector into a U-Net module of the image generation model, and obtaining updated potential image feature and key point condition vector; Inputting the updated latent image features and key point condition vector into a key point heat map extraction module based on condition guidance to obtain a key point heat map; The key point coordinates of the object are determined based on the key point heat map.

[0006] In an optional embodiment, it also includes: The condition-guided key point heat map extraction module is trained using the heat map expected loss function.

[0007] In an optional embodiment, the heat map expected loss function calculates the expected score between the estimated heat map and the true heat map by integration; The goal of the heat map expected loss function is to maximize the expected score.

[0008] In an optional embodiment, the image data to be recognized is input into an encoder of an image generation model to obtain potential image features, including: The image data to be recognized is input into the pre-trained VAE encoder in the image generation model to obtain the potential image features.

[0009] In an optional embodiment, initializing the key point condition vector includes: Get the number of key points to be located; Based on the number of key points to be located, the key point condition vectors are initialized, wherein each key point corresponds to a condition vector.

[0010] In an optional embodiment, the updated potential image features and key point condition vectors are input into a key point heat map extraction module based on condition guidance to obtain a key point heat map, including: Inputting the updated latent image features and key point condition vector into a condition-guided key point heat map extraction module; The condition-guided key point heat map extraction module is used to calculate the cosine similarity between the updated key point condition vector and the potential image feature, and normalize it using the softmax function to obtain the heat map of each key point.

[0011] In an optional embodiment, determining the key point coordinates of the object based on the key point heat map includes: Get the heat map corresponding to each key point; In each heat map channel, the coordinates of the key point are determined based on the position of the maximum pixel value.

[0012] According to another aspect of the present application, there is provided a device for locating key points of an object, comprising: An input module, used for inputting the image data to be recognized into the encoder of the image generation model to obtain potential image features; A processing module, used for initializing a key point condition vector, inputting the potential image features and the key point condition vector into a U-Net module of the image generation model, and obtaining updated potential image features and key point condition vector; A heat map extraction module, used for inputting the updated potential image features and key point condition vectors into a key point heat map extraction module based on condition guidance to obtain a key point heat map; A positioning module is used to determine the key point coordinates of the object based on the key point heat map.

[0013] In an optional embodiment, it also includes: A training module is used to train the condition-guided key point heat map extraction module using a heat map expected loss function.

[0014] According to another aspect of the embodiments of the present application, there is also provided an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the object key point positioning method through the computer program.

[0015] The technical solution provided by the embodiments of the present application may have the following beneficial effects: The entire key point positioning model of this application is built based on a pre-trained image generation model. The image generation model is trained on large-scale image data and learns a wealth of prior knowledge, including the structure, texture, semantic information, etc. of the image. Through the encoder and U-Net, the model can convert the input image into a universal feature representation and use these feature representations to locate key points. This universal feature representation is based on learning from a large amount of image data, so that the model can adapt to different image fields and object categories. The model can generate specific key point distribution maps according to different conditions through a conditional guidance mechanism. This enables the model to flexibly locate the key points of different types of objects. Therefore, the method of this application can improve the generalization ability and positioning accuracy of the model.

[0016] Furthermore, this application also proposes to train the model based on the heat map expected loss to improve the accuracy of key point positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 is a flow chart of an optional object key point positioning method according to an embodiment of the present application; Figure 2 is a schematic diagram of a key point positioning model according to an embodiment of the present application; Figure 3 is a schematic diagram of heat map extraction according to an embodiment of the present application; Figure 4 is a structural schematic diagram of an object key point positioning device according to an embodiment of the present application; Figure 5 It is a schematic diagram of the structure of an optional electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0019] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0020] Generic object keypoint localization is a fundamental computer vision task for understanding object structures. Existing keypoint localization methods suffer from poor performance and generalization due to their limited training data that cannot provide generic shape and semantic cues. To address this challenge, this application does not rely on large-scale training data, but instead leverages the rich prior knowledge in large generative models. An efficient generic localization method is proposed. Generative priors are extracted from a pre-trained image generative model by computing the correlation graph between image latent features and conditional embeddings. These priors are then optimized using the proposed heatmap expectation loss to achieve object keypoint localization. Benefiting from the rich knowledge of generative priors in understanding object semantics and structure, superior performance is achieved on various object keypoint localization benchmarks. It demonstrates more significant performance gains in cross-domain, few-shot, and zero-shot evaluation settings.

[0021] The following is combined with Figure 1-3The object key point positioning method of the embodiment of the present application is described in detail. Figure 1 As shown, the method mainly includes the following steps: S101 inputs the image data to be recognized into the encoder of the image generation model to obtain potential image features.

[0022] In one embodiment, the image data to be recognized is input into a pre-trained VAE encoder in an image generation model to obtain potential image features. The image data may be a video frame or a selected picture.

[0023] Specifically, the object structure can be represented by a set of key points To represent, where K i represents the coordinates of the i-th type of key point, such as the shoulder of a human body or the base of an animal's tail, and n is the total number of key point types defined for the object. Therefore, the goal of object key point localization is to estimate the coordinates of the target key point from the input image.

[0024] Typical object keypoint localization methods learn the model Φ(·) entirely from training data, i.e., train on a specific domain. This process makes them limited to specific image domains or object categories and leads to poor performance and generalization. Instead of learning Φ(·) from a specific domain, this application aims to leverage the rich prior knowledge in image generation models for localization. For example, the latent diffusion model LDM trained on large-scale image data has shown encouraging performance in understanding scene structure and has been successfully applied to tasks such as depth estimation or edge detection. This motivates us to learn a general localization model Φ(·) by leveraging the generative prior knowledge of LDM.

[0025] The image generation model includes a pre-trained VAE encoder. First, the image generation model LDM first uses the pre-trained VAE encoder E to encode the input image I into a latent image feature z=E(I).

[0026] S102 initializes the key point condition vector, inputs the latent image features and the key point condition vector into the U-Net module of the image generation model, and obtains updated latent image features and the key point condition vector.

[0027] Further, the number of key points to be located is obtained; based on the number of key points to be located, the key point condition vectors are initialized, wherein each key point corresponds to a condition vector. These condition vectors are used to guide the model to locate specific key points, such as shoulders and elbows of the human body.

[0028] Furthermore, the latent image features and key point condition vectors are input into the U-Net module of the image generation model to obtain updated latent image features and key point condition vectors.

[0029] LDM is generated in the latent space, that is, a denoising U-Net is trained. The U-Net receives the latent image features and the key point condition vector, and the U-Net outputs the updated latent image features and the key point condition vector. The U-Net updates the latent features through the denoising process to make them closer to the target distribution, and updates the condition vector to better adapt to the key point localization task.

[0030] S103 inputs the updated latent image features and key point condition vectors into a key point heat map extraction module based on condition guidance to obtain a key point heat map.

[0031] In an optional embodiment, the updated potential image features and key point condition vectors are input into a key point heat map extraction module based on condition guidance to obtain a key point heat map, including: The updated latent image features and key point conditional vectors are input into the condition-guided key point heat map extraction module; the condition-guided key point heat map extraction module is used to calculate the cosine similarity between the updated key point conditional vector and the latent image features, and normalize them using the softmax function to obtain the heat map of each key point.

[0032] In the implementation of the present application, LDM first uses a pre-trained VAE encoder E to encode the image I into a potential image feature z = E (I), and the decoder D can reconstruct the original data from z, that is, I = D (z) ≈ I. LDM is generated in the latent space, that is, a denoising U-Net is trained, U is used to predict the current noise input z t , the noise between time step t and condition c.

[0033] LDM fixes E(·) and only minimizes the difference between the predicted noise and its true value distance to train U(·). Therefore, the generated prior is embedded in U(·). However, how to extract such information from LDM remains a problem. Previous work has shown that directly fine-tuning U(·) on a specific task can achieve satisfactory results, but it is not easy to directly apply this approach to keypoint localization. Keypoint localization requires outputting multiple types of keypoints for a single object, the number of which is greater than the number of image channels, which makes it infeasible to use a decoder D(·) to recover the latent feature z into an image-level representation.

[0034] Therefore, this application proposes a condition-guided heatmap extraction module to extract the generative prior in U(·) to achieve Φ(·) and perform universal keypoint localization. Instead of fine-tuning U(·) to output the corresponding heatmap, the correlation map C between the latent feature map z and the condition ci is used to generate the final heatmap, which is achieved through the proposed condition-guided keypoint heatmap extraction module (CKHE), that is, in this way, the keypoint type is encoded into the condition ci, allowing us to flexibly locate any type of keypoint.

[0035] Condition-guided keypoint heatmap extraction. Different from previous works that directly fine-tune U-Net to estimate heatmaps, we use the controllable generation mechanism in LDM to generate heatmaps. Most current generative models implement controllable mechanisms through conditional inputs such as text or images. The conditions are injected into the generation process through the attention mechanism widely adopted in Transformer. Specifically, given the latent features z∈Rm×d and the conditional embedding c∈Rn×d, we pass them to U(·) to obtain the updated latent features and conditions. The cross attention can be calculated as follows: (1) Among them, S∈Rm×n represents the attention score between the latent feature z and the condition c, m and n represent the number of image tags and keypoint condition embeddings, respectively, and d represents the dimension of the feature. It should be noted that the softmax operation is performed along different keypoint conditions, which indicates that each tag from the latent feature z independently finds similar keypoint conditions and aggregates their features.

[0036] Although S in the above formula (1) can be interpreted as a heat map for localization, there are several defects in directly adopting it. First, the similarity calculation in S is independent between the m image regions, which does not conform to the property that the heat map has a high response only at the target spatial location, resulting in inaccurate key point distribution estimation. Second, the different conditions in S are correlated, and it is difficult to distinguish spatially adjacent key points because each image region can only produce a high response to one type of key point. Therefore, we propose a condition-guided key point heat map extraction to avoid the above limitations. Unlike the cross attention in the above formula, the key point heat map extraction CKHE considers calculating the correlation map between the latent feature z and the key point condition c, so that each marker in z is correlated, but the different conditions in c are independent. To achieve this goal, we exchange the calculation order in the above formula and calculate the softmax on the latent image region instead of the key point condition, that is: (2) Where C∈Rn×m represents the heatmap estimated by extracting CKHE from the keypoint heatmap. In addition to applying formula (2) of this application, we also need to fine-tune the corresponding attention layer parameters to adapt to the changes in calculation.

[0037] like Figure 3 As shown in the figure, different interaction mechanisms between image latent features and keypoint conditions are demonstrated. (a) LDM uses a cross-attention mechanism to aggregate conditional information into latent features, which is not suitable for keypoint localization and leads to poor heatmap results. (b) Our proposed CKHE model models the keypoint distribution on the image latent features and decouples the keypoint conditions, thereby generating a more localized heatmap and distinguishing spatially adjacent keypoints.

[0038] Note that although Equation (1) and Equation (2) are similar, they differ in many aspects. As mentioned above, Equation (2) uses keypoint conditions as queries to compute similarity scores, so the tokens in the latent z are naturally correlated, i.e., only a small number of tokens have high probability in the estimated correlation map C. In addition, Equation (2) also removes the correlation of keypoint conditions, thereby enabling the detection of spatially adjacent keypoints. Figure 3 As shown, CKHE is able to model keypoint distribution more accurately and generate more localized heatmaps, thus achieving better performance.

[0039] S104 determines the key point coordinates of the object based on the key point heat map.

[0040] In an optional embodiment, determining the key point coordinates of the object based on the key point heat map includes: obtaining the heat map corresponding to each key point; and determining the coordinates of the key point based on the position with the maximum pixel value in each heat map channel.

[0041] In an optional embodiment, it also includes: using a heat map expected loss function to train a condition-guided key point heat map extraction module. The heat map expected loss function calculates the expected score between the estimated heat map and the real heat map by integration; the goal of the heat map expected loss function is to maximize the expected score.

[0042] Our goal is to optimize the condition c so that the obtained correlation map C can accurately locate the target key points. To achieve this goal, we measure C and the target true value The loss is calculated by taking the difference between Distance, that is:

[0043] However, adding element-wise supervision to the above formula is invalid because is a manually defined non-normalized Gaussian distribution with fixed variance, which cannot reflect the underlying true distribution and will prevent the model from learning an accurate distribution to simulate the target keypoints. Although the underlying distribution of the target keypoints is unknown, it should be consistent with There is a large overlap to produce a high response at the target location. Therefore, this application turns to simulating C and , and proposes a heatmap expectation loss (HEL) to address the above limitations. Points target To calculate the expected score to simulate C and The goal of HEL is to maximize the expected score, that is:

[0044] Where m represents the number of image tags. A higher expected score indicates and The greater the overlap between Figure 3 The heatmap in is clearer.

[0045] With L MSE In comparison, L HEL It can fully utilize the normalization properties of C to generate more effective gradients during training, thus achieving better performance.

[0046] Compared with the traditional mean square error (L MSE ), HEL has the following advantages: More efficient gradients: HEL computes the expected score by integrating, making full use of the estimated distribution The normalization property of HEL generates more effective gradients during training. Better performance: By maximizing the expected score, HEL is able to more accurately optimize keypoint localization, resulting in better performance. Faster convergence: The gradients generated by HEL are more effective in the early stages of training, helping the model converge faster.

[0047] In order to facilitate understanding of the method of the embodiment of the present application, Figure 2 Further description.

[0048] like Figure 2As shown in the figure, the object key point localization model includes a pre-trained image generation model, a condition-guided key point heat map extraction module, and a heat map expectation loss. Among them, the image generation model adopts the latent diffusion model LDM, including an encoder and a denoising U-Net. The object image is taken as input, and a set of key point conditions are initialized to extract the generation prior. The image input is processed by a frozen VAE encoder to obtain the latent image features. The denoising U-Net takes the latent features and key point conditions as input and generates the corresponding updated output features. The latent features and key point conditions are passed to the proposed CKHE module to obtain the key point distribution map. The whole process is optimized in an end-to-end manner through the proposed heat map expectation loss.

[0049] This application proposes GenLoc, a generalized object keypoint localization method. GenLoc exploits generative priors in image generation models to perform keypoint localization. It first adopts the proposed CKHE to extract the correlation map between conditional and latent image features, and then optimizes the correlation map through our heatmap expectation loss to perform localization. Experiments on various keypoint localization benchmarks show that the proposed method has strong generalization ability.

[0050] According to another aspect of the embodiments of the present application, a device for locating key points of an object for implementing the above-mentioned method for locating key points of an object is also provided. Figure 4 As shown, the device comprises: An input module 401 is used to input the image data to be recognized into the encoder of the image generation model to obtain potential image features; Processing module 402, used to initialize the key point condition vector, input the potential image features and the key point condition vector into the U-Net module of the image generation model, and obtain updated potential image features and the key point condition vector; The heat map extraction module 403 is used to input the updated potential image features and key point condition vectors into the key point heat map extraction module based on condition guidance to obtain the key point heat map; The positioning module 404 is used to determine the key point coordinates of the object based on the key point heat map.

[0051] In an optional embodiment, it also includes: The training module is used to train the condition-guided key point heat map extraction module using the heat map expected loss function.

[0052] It should be noted that the object key point positioning device provided in the above embodiment only uses the division of the above functional modules as an example when executing the object key point positioning method. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the object key point positioning device provided in the above embodiment and the object key point positioning method embodiment belong to the same concept, and the implementation process thereof is detailed in the method embodiment, which will not be repeated here.

[0053] According to another aspect of the embodiments of the present application, an electronic device corresponding to the object key point positioning method provided in the above embodiments is also provided to execute the above object key point positioning method.

[0054] Please refer to Figure 5 , which shows a schematic diagram of an electronic device provided by some embodiments of the present application. Figure 5 As shown, the electronic device includes: a processor 500, a memory 501, a bus 502 and a communication interface 503, and the processor 500, the communication interface 503 and the memory 501 are connected via the bus 502; the memory 501 stores a computer program that can be run on the processor 500, and when the processor 500 runs the computer program, it executes the object key point positioning method provided in any of the aforementioned embodiments of the present application.

[0055] The memory 501 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk storage. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 503 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0056] The bus 502 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. The memory 501 is used to store programs, and the processor 500 executes the program after receiving the execution instruction. The object key point positioning method disclosed in any implementation of the embodiment of the present application may be applied to the processor 500, or implemented by the processor 500.

[0057] The processor 500 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 500. The above processor 500 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly embodied as a hardware decoding processor to be executed, or the hardware and software modules in the decoding processor can be executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 501, and the processor 500 reads the information in the memory 501 and completes the steps of the above method in combination with its hardware.

[0058] The electronic device provided in the embodiment of the present application and the object key point positioning method provided in the embodiment of the present application are based on the same inventive concept and have the same beneficial effects as the methods adopted, operated or implemented therein.

[0059] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0060] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the attached claims.

Claims

1. A method for locating key points of an object, characterized in that: include: Input the image data to be recognized into the encoder of the image generation model to obtain the potential image features; Initializing a key point condition vector, inputting the potential image feature and the key point condition vector into a U-Net module of the image generation model, and obtaining updated potential image feature and key point condition vector; Inputting the updated latent image features and key point condition vector into a key point heat map extraction module based on condition guidance to obtain a key point heat map; The key point coordinates of the object are determined based on the key point heat map.

2. The method according to claim 1, characterized in that Also includes: The condition-guided key point heat map extraction module is trained using the heat map expected loss function.

3. The method according to claim 2, characterized in that The heat map expected loss function calculates the expected score between the estimated heat map and the true heat map by integration; The goal of the heat map expected loss function is to maximize the expected score.

4. The method according to claim 1, characterized in that: The image data to be recognized is input into the encoder of the image generation model to obtain potential image features, including: The image data to be recognized is input into the pre-trained VAE encoder in the image generation model to obtain the potential image features.

5. The method according to claim 1, characterized in that Initialize the key point condition vector, including: Get the number of key points to be located; Based on the number of key points to be located, the key point condition vectors are initialized, wherein each key point corresponds to a condition vector.

6. The method according to claim 1, characterized in that The updated latent image features and key point condition vectors are input into a key point heat map extraction module based on condition guidance to obtain a key point heat map, including: Inputting the updated latent image features and key point condition vector into a condition-guided key point heat map extraction module; The condition-guided key point heat map extraction module is used to calculate the cosine similarity between the updated key point condition vector and the potential image feature, and normalize it using the softmax function to obtain the heat map of each key point.

7. The method according to claim 1, characterized in that Determining the key point coordinates of the object based on the key point heat map includes: Get the heat map corresponding to each key point; In each heat map channel, the coordinates of the key point are determined based on the position of the maximum pixel value.

8. An object key point positioning device, characterized in that: include: An input module, used for inputting the image data to be recognized into the encoder of the image generation model to obtain potential image features; A processing module, used for initializing a key point condition vector, inputting the potential image features and the key point condition vector into a U-Net module of the image generation model, and obtaining updated potential image features and key point condition vector; A heat map extraction module, used for inputting the updated potential image features and key point condition vectors into a key point heat map extraction module based on condition guidance to obtain a key point heat map; A positioning module is used to determine the key point coordinates of the object based on the key point heat map.

9. The device according to claim 8, characterized in that Also includes: A training module is used to train the condition-guided key point heat map extraction module using a heat map expected loss function.

10. An electronic device, characterized in that: The invention comprises a processor and a memory storing program instructions, wherein the processor is configured to execute the object key point positioning method according to any one of claims 1 to 7 when executing the program instructions.

Images