Feature labeling model training method, feature labeling method, device and medium

Abstract

The invention relates to the technical field of computers, in particular to a training method of a feature annotation model, a feature annotation method, equipment and a medium, and aims to solve the technical problems that an existing video annotation result is poor in time sequence understanding ability, single in task structure, limited in data format and high in annotation cost. In order to achieve the purpose, labeling of a preset type of labeling target is carried out on a training data set used for training, and a structured labeling result of each video frame sequence of the training data set is obtained. And training the feature labeling model according to the labeled training data set so as to obtain the trained feature labeling model, so that the detection, tracking and feature description capabilities of the feature labeling model can be effectively trained, and the feature labeling model can cope with more complex feature labeling tasks. And continuous target tracking, target identity consistency maintenance and cross-frame feature induction capability of the feature labeling model are realized.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a training method, feature annotation method, device, and medium for a feature annotation model. Background Technology

[0002] Existing multimodal large models (such as GLM-4.5v, Qwen2.5-VL series, etc.) perform well in processing static image understanding tasks, but they have obvious limitations when processing tasks involving temporal information, mainly in the following aspects: 1. Lack of temporal coherence understanding: Existing models mostly analyze videos frame by frame independently. The models lack an effective mechanism to associate the same target in different frames, i.e., to perform temporal correlation, which makes it impossible to perform stable tracking and understand the behavioral evolution of the target.

[0003] 2. Limited Task Definition: Traditional fine-tuning methods are typically designed for a single task (such as detection or recognition only), lacking a unified framework to integrate multiple tasks such as detection, tracking, and feature description for end-to-end learning. This results in fragmented model capabilities, making it difficult to accomplish complex tasks such as "inputting a video and outputting temporal IDs and feature descriptions."

[0004] 3. Data Format Limitations: Commonly used object detection or image description datasets use frame-by-frame annotation formats. For example, bounding boxes of the same person in different frames are independent anonymous annotations, lacking a unique ID across the video and failing to provide a unified feature summary across frames. This format makes it impossible to train models to form temporal identity consensus and feature induction capabilities.

[0005] 4. High cost of data annotation: Building a high-quality video temporal annotation dataset requires a lot of manpower to annotate, track and assign IDs frame by frame, which is time-consuming, labor-intensive and prone to errors, becoming a bottleneck for technology implementation.

[0006] Accordingly, a new feature annotation scheme is needed in this field to solve the above problems. Summary of the Invention

[0007] In order to overcome the above-mentioned shortcomings, this application is proposed to solve, or at least partially solve, the technical problems of how to effectively address the poor temporal understanding ability, simple task structure, limited data format, and high annotation cost of existing video annotation results.

[0008] In a first aspect, a method for training a feature annotation model is provided, characterized in that the method includes: The training dataset used for training is labeled with a preset type of labeled target, and the structured labeling result of each video frame sequence in the training dataset is obtained; the structured labeling result includes the target ID and spatial information of the labeled target in each video frame in the video frame sequence, as well as the global inductive feature of the labeled target in all video frames in the video frame sequence; the target ID of the same labeled target has global consistency in each video frame in the video frame sequence; The feature annotation model is trained based on the labeled training dataset to obtain the trained feature annotation model.

[0009] In one technical solution of the above-mentioned feature annotation model training method, the step of annotating the training dataset used for training with a preset type of annotation target to obtain the structured annotation result of each video frame sequence in the training dataset includes: For each video frame sequence in the training dataset, the video frame sequence is labeled with the labeled target based on a preset first labeling model, and the model labeling result is obtained; Based on the model annotation results, obtain the structured annotation results.

[0010] In one technical solution of the above-mentioned feature annotation model training method, the step of annotating the video frame sequence with the annotation target based on a preset first annotation model to obtain the model annotation result includes: Based on the first annotation model and the preset first annotation prompt words, the video frame sequence is annotated with the annotation target to obtain the model annotation result; The first prompt word is a prompt text that instructs the first annotation model to generate model annotation results including the target ID and spatial information of the labeled target in each video frame of the video frame sequence, as well as the global inductive features of the labeled target in all video frames of the video frame sequence.

[0011] In one technical solution of the above-mentioned feature annotation model training method, obtaining the structured annotation result based on the model annotation result includes: The model annotation results are manually verified and corrected to obtain the structured annotation results.

[0012] In one technical solution of the above-mentioned feature annotation model training method, the step of training the feature annotation model based on the annotated training dataset to obtain the trained feature annotation model includes: The video frame sequence containing the structured annotation results in the training dataset is input into the feature annotation model to train the feature annotation model and obtain the trained feature annotation model.

[0013] In one technical solution of the above-mentioned feature annotation model training method, the feature annotation model is a second large annotation model; The step of training the feature annotation model to obtain the trained feature annotation model includes: Set a preset second prompt word; the second prompt word is a prompt text that instructs the second annotation model to achieve the annotation and temporal correlation of the preset type of annotation target; Based on the second prompt word and the video frame sequence containing the structured annotation results in the training dataset, the second annotation model is trained to obtain the trained second annotation model.

[0014] In a second aspect, a feature annotation method is provided, the method comprising: Based on the trained feature annotation model, the passenger video frame sequence is annotated to obtain the feature annotation results of the passenger video frame sequence. The feature annotation model is obtained by training according to any one of the feature annotation model training methods described above.

[0015] In a third aspect, an electronic device is provided, comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program, which, when executed by the at least one processor, implements the method described in any of the above-described technical solutions for the feature annotation model training method.

[0016] In a fourth aspect, an electronic device is provided, comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program, which, when executed by the at least one processor, implements the method described in any of the above-described feature annotation methods.

[0017] In a fifth aspect, a computer-readable storage medium is provided, wherein a plurality of program codes are stored therein, the program codes being adapted to be loaded and run by a processor to perform the technical solution of the above-described feature annotation model training method or any of the technical solutions of the above-described feature annotation method.

[0018] The above-described technical solutions of this application have at least one or more of the following beneficial effects: In implementing the feature annotation model training method provided in this application, the training dataset used for training is annotated with pre-defined types of target annotations to obtain structured annotation results for each video frame sequence in the training dataset. The structured annotation results include the target ID and spatial information of the labeled target in each video frame sequence, as well as the globally inductive features of the labeled target across all video frames in the video frame sequence. Furthermore, the target ID of the same labeled target has global consistency across all video frames in the video frame sequence. Based on the annotated training dataset, the feature annotation model is trained to obtain a well-trained feature annotation model. Through the above configuration, this application enables the feature annotation model to possess temporal understanding capabilities between video frame sequences based on structured annotation results. This allows for stable tracking of annotated targets within the video frame sequence and an understanding of the behavioral evolution of the annotated targets. Simultaneously, it effectively trains the feature annotation model's detection, tracking, and feature description capabilities, enabling it to handle more complex feature annotation tasks. Furthermore, training the feature annotation model with a training dataset containing structured annotation results allows it to learn a unified data annotation format, achieving continuous target tracking, consistent target identity maintenance, and cross-frame feature induction capabilities. Training based on target IDs enables effective temporal supervision of the target ID during training, resulting in a more robust feature annotation model to scenarios such as occlusion and deformation, effectively reducing target ID switching errors during target tracking. Consequently, feature annotation based on the trained feature annotation model yields higher-quality feature annotation results, eliminating the need for manual adjustments when applying the feature annotation model and effectively reducing feature annotation costs. Attached Figure Description

[0019] The disclosure of this application will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this application. Wherein: Figure 1 This is a schematic diagram of the main steps of a feature annotation model training method according to an embodiment of this application; Figure 2 This is a structured annotation result based on an example of an embodiment of this application; Figure 3 This is a schematic diagram of the main steps in obtaining a training dataset according to one embodiment of the present application. Figure 4 This is a schematic flowchart of the main steps of a feature annotation method according to an embodiment of this application; Figure 5This is a schematic flowchart of the main steps of a feature annotation method according to one embodiment of this application. Detailed Implementation

[0020] Some embodiments of this application are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of this application and are not intended to limit the scope of protection of this application.

[0021] In the description of this application, "module" and "processor" can include hardware, software, or a combination of both. A module can include hardware circuitry, various suitable sensors, communication ports, and memory, and may also include software components, such as program code, or a combination of software and hardware. The term "A and / or B" means all possible combinations of A and B, such as only A, only B, or A and B. The terms "at least one A or B" or "at least one of A and B" have a similar meaning to "A and / or B" and can include only A, only B, or A and B. The singular forms of the terms "a" and "this" can also include plural forms.

[0022] See appendix Figure 1 , Figure 1 This is a schematic diagram illustrating the main steps of a feature annotation model training method according to an embodiment of this application. Figure 1 As shown, the feature annotation model training method in this application embodiment mainly includes the following steps S101 to S102.

[0023] Step S101: Label the training dataset used for training with pre-defined types of labeled targets, and obtain the structured labeling results of each video frame sequence in the training dataset; the structured labeling results include the target ID and spatial information of the labeled target in each video frame in the video frame sequence, as well as the global inductive features of the labeled target in all video frames in the video frame sequence; the target ID of the same labeled target in each video frame in the video frame sequence has global consistency.

[0024] In this embodiment, the dataset used for training can be pre-labeled with target information to obtain structured labeling results for each video frame sequence.

[0025] In one implementation, the preset target can be a passenger, pedestrian, or other target in the video frame sequence.

[0026] In one embodiment, step S101 may include steps S1011 and S1012: Step S1011: For each video frame sequence in the training dataset, label the target of the video frame sequence based on the preset first labeling model, and obtain the model labeling result.

[0027] In this embodiment, the first annotation model can be applied to annotate the video frame sequences in the training dataset with the annotation targets.

[0028] In one implementation, the first labeled model can be a multimodal large model with grounding capabilities, such as GLM-4.5v or Qwen2.5-VL.

[0029] In one implementation, a first annotation prompt word can be set. This first annotation prompt word is a prompt text instructing the first annotation model to generate model annotation results that include the target ID and spatial information of the target in each video frame of the video frame sequence, as well as the globally inductive features of the target across all video frames in the video frame sequence. That is, the core instructions of the first annotation prompt word need to include: Task definition: Clearly define the model annotation target and annotation task (e.g., analyzing continuous images, identifying passengers getting off the bus, etc.); Core rules: Emphasize the continuity of temporal target IDs in labeled targets, in-vehicle passenger screening, and feature description dimensions; Output constraints: Force the model to output in a specified structured annotation format (e.g., JSON format), constraining the model's free output to standardized structured annotation results.

[0030] The process of obtaining model annotation results based on the first annotation model and the first annotation prompt words utilizes the powerful zero-shot understanding capability of the first annotation model (such as GLM-4.5v, Qwen2.5-VL, etc.) to quickly obtain model annotation results containing the detection box of the labeled target, the temporal target ID, and the feature description (including global inductive features).

[0031] Step S1012: Obtain the structured annotation results based on the model annotation results.

[0032] In this embodiment, the model annotation results can be manually verified and corrected to obtain structured annotation results. The manual verification and correction may include the following: Correct the incorrect detection box (bbox); Correcting incorrect target ID associations (ID Switch); Correct or refine inaccurate feature descriptions; Supplementing missing passengers, etc.

[0033] Because the first annotation model has a strong zero-shot understanding capability, the manual verification and correction process only needs to correct a small amount of content. While ensuring the accuracy of the structured annotation results, it can also effectively improve the annotation efficiency of the training dataset, freeing people from the heavy annotation work and allowing them to carry out more efficient manual verification and correction work.

[0034] In one implementation, the structured annotation results can be in JSON (JavaScript Object Notation) format or a similar structured data format. See the appendix for details. Figure 2 , Figure 2 The example shown illustrates a structured annotation result, where the information of a video frame sequence is organized into a list. Each element in this list represents information about a single video frame or a key video frame, and the last element is a global inductive feature summarizing the entire video frame sequence. That is, the same target ID (e.g., P001, P002) remains unique and constant across different video frames. The final `passenger_features` field is a feature summary of the passenger corresponding to each target ID in the entire video frame sequence, i.e., a global inductive feature summarizing feature.

[0035] Figure 2 The structured annotation results shown in the figure remain unchanged across different video frames by using the target ID. This allows the feature annotation model to learn that the labeled target corresponding to the target ID is the same labeled target (e.g., the same person), effectively reflecting the temporal identity consistency of labeled targets in the video frame sequence. At the same time, the spatial information (bbox) and target ID in each video frame are bound together, which can provide joint supervision signals for the feature annotation model and realize spatiotemporal joint annotation of the structured annotation results. Furthermore, the final global inductive features (passenger_features) ensure that the feature annotation model not only sees local features during training, but also summarizes and integrates visual information across time to form a stable and comprehensive understanding of each labeled target, avoiding the problem of inconsistent descriptions frame by frame.

[0036] Step S102: Train the feature annotation model based on the labeled training dataset to obtain the trained feature annotation model.

[0037] In this embodiment, a sequence of video frames containing structured annotation results in the training dataset can be input into the feature annotation model to train the feature annotation model and obtain the trained feature annotation model.

[0038] In one implementation, the feature annotation model can be a second large annotation model, such as the GLM-4.5v or Qwen2.5VL series models. The second large annotation model and the first annotation model can be the same model or different models.

[0039] In one embodiment, step S102 may further include steps S1021 and S1022: Step S1021: Set a preset second prompt word; the second prompt word is a prompt text that instructs the second annotation model to achieve the annotation and temporal association of the preset type of annotation target.

[0040] In this embodiment, a second prompt word containing explicit instructions can be set. The second prompt word can contain a complete and clear task description, clearly indicate the task type of the model, and clearly define the workflow for the model, such as detecting labeled targets and realizing temporal correlation.

[0041] Step S1022: Based on the second prompt word and the video frame sequence containing structured annotation results in the training dataset, train the second annotation model to obtain the trained second annotation model.

[0042] In this embodiment, a sequence of video frames containing structured annotations from the training dataset can be input into a second large-scale annotation model for training. The training objective is for the second large-scale annotation model to accurately generate or predict the global inductive features (i.e., passenger_features) of the annotated targets in a series of images (i.e., video frame sequences) and annotated target fields (e.g., the passengers field) after inputting these. In this way, the second large-scale annotation model can learn to detect and locate the annotated targets (e.g., passengers) in each video frame of the video frame sequence during training, obtaining the spatial information (e.g., bounding boxes) of the annotated targets; establish associations between the same annotated target (e.g., passengers) in different videos; and perform stable feature aggregation (i.e., obtain global inductive features) of the annotated targets in the video frame sequence.

[0043] In one implementation, see Appendix Figure 3 ,like Figure 3 As shown, the training dataset can be obtained by following these steps: Input an unlabeled video frame sequence into the first annotation model; based on the first annotation model and the first prompt word, annotate the labeled targets in the video frame sequence in the training dataset, and generate the model annotation results, as shown below. Figure 2The JSON data format shown is used for manual verification and correction of the model annotation results. This mainly includes modifying the bounding boxes (bboxes) for inaccurate model annotation results; correcting target ID association errors to ensure temporal consistency of the labeled targets; and refining / correcting feature descriptions for inaccurate model annotation results. Based on the manually verified and corrected model annotation results, structured annotation results are obtained, thereby obtaining a high-quality training dataset.

[0044] Based on the methods described in steps S101 to S103 above, this embodiment of the application annotates the training dataset used for training with preset types of labeled targets, obtaining structured annotation results for each video frame sequence of the training dataset. The structured annotation results include the target ID and spatial information of the labeled target in each video frame sequence, as well as the globally inductive features of the labeled target across all video frames in the video frame sequence, and the target ID of the same labeled target has global consistency across each video frame in the video frame sequence. Based on the annotated training dataset, the feature annotation model is trained to obtain a trained feature annotation model. Through the above configuration, this embodiment of the application enables the feature annotation model to possess temporal understanding capabilities between video frame sequences based on structured annotation results. This allows for stable tracking of annotated targets within the video frame sequence and an understanding of the behavioral evolution of the annotated targets. Simultaneously, it effectively trains the feature annotation model's detection, tracking, and feature description capabilities, enabling the model to handle more complex feature annotation tasks. Furthermore, training the feature annotation model with a training dataset containing structured annotation results allows it to learn a unified data annotation format, achieving continuous target tracking, consistent target identity maintenance, and cross-frame feature induction capabilities. Training based on target IDs enables effective temporal supervision of target IDs during training, resulting in a more robust feature annotation model to scenarios such as occlusion and deformation, effectively reducing target ID switching errors during target tracking. Therefore, feature annotation based on the trained feature annotation model yields higher-quality feature annotation results, eliminating the need for manual adjustments when applying the feature annotation model and effectively reducing feature annotation costs.

[0045] Furthermore, this application also provides a feature annotation method.

[0046] See appendix Figure 4 , Figure 4 This is a schematic flowchart illustrating the main steps of a feature annotation method according to an embodiment of this application. Figure 4 As shown, the feature annotation method in this embodiment mainly includes the following steps S201: Step S201: Based on the trained feature annotation model, the passenger video frame sequence is annotated to obtain the feature annotation results of the passenger video frame sequence; wherein, the feature annotation model is obtained by training according to the feature annotation model training method described in the above feature annotation model training method embodiment.

[0047] In this embodiment, the passenger video frame sequence can be labeled based on the trained feature labeling model to obtain the feature labeling results of the passenger video frames.

[0048] In one embodiment, the feature annotation model of this application can be a domain expert model used for bus passenger analysis, smart bus, passenger flow identification, and bus capacity optimization.

[0049] In one implementation, see Appendix Figure 5 ,like Figure 5 As shown, the feature annotation method in this application embodiment may include three steps: data preparation, model training and deployment, and application deployment. For data preparation, the original video to be labeled can be input and frame extraction can be performed (i.e., to obtain the video frame sequence in the training dataset); a human-computer collaborative data construction process can be carried out (i.e., the video sequence can be labeled by the first labeling model and manually verified and corrected) to obtain a time-series structured dataset (i.e., a training dataset containing structured labeling results).

[0050] For model training and deployment, multimodal large models (such as GLM-4.5v and Qwen2.5VL) can be fine-tuned based on time-series structured datasets to produce domain expert models (i.e., well-trained multimodal large models).

[0051] For application deployment, feature annotation can be performed on the input new video based on the domain expert model, thereby obtaining JSON results with time sequence ID, detection boxes and feature descriptions (i.e., feature annotation results).

[0052] It should be noted that although the steps in the above embodiments are described in a specific order, those skilled in the art will understand that in order to achieve the effect of this application, different steps do not necessarily have to be executed in such an order. They can be executed simultaneously (in parallel) or in other orders. These adjusted solutions are equivalent to the technical solutions described in this application and therefore will also fall within the protection scope of this application.

[0053] Those skilled in the art will understand that all or part of the processes in the method of the above-described embodiment can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include any entity or device capable of carrying the computer program code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0054] Another aspect of this application provides a computer-readable storage medium.

[0055] In one embodiment of a computer-readable storage medium according to this application, the computer-readable storage medium can be configured to store a program for performing the feature annotation model training method of the above-described method embodiments. This program can be loaded and run by a processor to implement the feature annotation model training method. For ease of explanation, only the parts related to the embodiments of this application are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of this application. The computer-readable storage medium can be a storage device formed by various electronic devices, such as a magnetic disk, hard disk, optical disk, flash memory, read-only memory, random access memory, etc. Optionally, in the embodiments of this application, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0056] Another aspect of this application provides an alternative computer-readable storage medium.

[0057] In one embodiment of a computer-readable storage medium according to this application, the computer-readable storage medium can be configured to store a program that performs the feature annotation method of the above-described method embodiments. This program can be loaded and run by a processor to implement the feature annotation method. For ease of explanation, only the parts related to the embodiments of this application are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of this application. The computer-readable storage medium can be a storage device formed by various electronic devices, such as a magnetic disk, hard disk, optical disk, flash memory, read-only memory, random access memory, etc. Optionally, in the embodiments of this application, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0058] Another aspect of this application provides an electronic device.

[0059] In an embodiment of an electronic device according to this application, the electronic device may include at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program that, when executed by the at least one processor, implements the method described in any of the embodiments of the above-described feature annotation model training method.

[0060] Another aspect of this application provides another electronic device.

[0061] In an embodiment of an electronic device according to this application, the electronic device may include at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program that, when executed by the at least one processor, implements the method described in any of the above-described embodiments of the feature annotation method.

[0062] In some embodiments of this application, the processor may be a central processing unit, a microprocessor, a graphics processor, a digital signal processor, or any other suitable processor. The processor has data and / or signal processing capabilities. The processor may be implemented in software, in hardware, or a combination of both.

[0063] The technical solution of this application has been described above with reference to one embodiment shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.

Claims

1. A method for training a feature annotation model, characterized in that, The method includes: The training dataset used for training is labeled with a preset type of labeled target, and the structured labeling result of each video frame sequence in the training dataset is obtained; the structured labeling result includes the target ID and spatial information of the labeled target in each video frame in the video frame sequence, as well as the global inductive feature of the labeled target in all video frames in the video frame sequence; the target ID of the same labeled target has global consistency in each video frame in the video frame sequence; The feature annotation model is trained based on the labeled training dataset to obtain the trained feature annotation model.

2. The feature annotation model training method according to claim 1, characterized in that, The step of labeling the training dataset for training with preset types of annotation targets and obtaining the structured annotation results for each video frame sequence in the training dataset includes: For each video frame sequence in the training dataset, the video frame sequence is labeled with the labeled target based on a preset first labeling model, and the model labeling result is obtained; Based on the model annotation results, obtain the structured annotation results.

3. The feature annotation model training method according to claim 2, characterized in that, The step of annotating the video frame sequence with the target based on a preset first annotation model and obtaining the model annotation result includes: Based on the first annotation model and the preset first annotation prompt words, the video frame sequence is annotated with the annotation target to obtain the model annotation result; The first prompt word is a prompt text that instructs the first annotation model to generate model annotation results including the target ID and spatial information of the labeled target in each video frame of the video frame sequence, as well as the global inductive features of the labeled target in all video frames of the video frame sequence.

4. The feature annotation model training method according to claim 2, characterized in that, The step of obtaining the structured annotation result based on the model annotation result includes: The model annotation results are manually verified and corrected to obtain the structured annotation results.

5. The feature annotation model training method according to claim 1, characterized in that, The step of training the feature annotation model based on the labeled training dataset to obtain the trained feature annotation model includes: The video frame sequence containing the structured annotation results in the training dataset is input into the feature annotation model to train the feature annotation model and obtain the trained feature annotation model.

6. The feature annotation model training method according to claim 5, characterized in that, The feature annotation model is the second large annotation model; The step of training the feature annotation model to obtain the trained feature annotation model includes: Set a preset second prompt word; the second prompt word is a prompt text that instructs the second annotation model to achieve the annotation and temporal correlation of the preset type of annotation target; Based on the second prompt word and the video frame sequence containing the structured annotation results in the training dataset, the second annotation model is trained to obtain the trained second annotation model.

7. A feature annotation method, characterized in that, The method includes: Based on the trained feature annotation model, the passenger video frame sequence is annotated to obtain the feature annotation results of the passenger video frame sequence. The feature annotation model is obtained by training according to the feature annotation model training method according to any one of claims 1 to 6.

8. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores a computer program, which, when executed by the at least one processor, implements the feature annotation model training method according to any one of claims 1 to 6.

9. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores a computer program, which, when executed by the at least one processor, implements the feature annotation method of claim 7.

10. A computer-readable storage medium storing a plurality of program codes, characterized in that, The program code is adapted to be loaded and run by a processor to perform the feature annotation model training method of any one of claims 1 to 6 or the feature annotation method of claim 7.

Images