Methods of generating a scene description, methods and systems of active question answering, and apparatuses
By performing target recognition and anchor point graph construction on video frames, scene descriptions are generated and proactive question answering is performed, solving the description and question answering tasks in complex scenes and achieving accurate and efficient description and question answering.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- IFLYTEK SOUTH CHINA ARTIFICIAL INTELLIGENCE RES INST GUANGZHOU CO LTD
- Filing Date
- 2022-06-15
- Publication Date
- 2026-05-19
AI Technical Summary
There is a lack of mature technical solutions in the current technology that can simultaneously solve scene description and visual question answering tasks in complex scenarios.
By performing target recognition on video frames, an anchor point map is constructed, and relevant descriptions are generated based on selection probabilities. This is combined with user questions to provide proactive question and answer.
It improves the accuracy and efficiency of scene description, avoids the generation of redundant descriptions, and can quickly answer user questions.
Smart Images
Figure CN115186131B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of information processing technology, and in particular to a method for generating scene descriptions, a method, system, and apparatus for proactive question answering. Background Technology
[0002] Reasoning tasks in complex scenarios can be divided into scene description and visual question answering. Scene description technology can be used for automatic movie subtitles and to help visually impaired people quickly understand their surroundings, while visual question answering technology helps humans explore unknown environments in an interactive way and can be used for visual navigation and chatbots. However, there is currently no mature technical solution that can simultaneously solve both scene description and visual question answering tasks. Summary of the Invention
[0003] The main technical problem addressed by this application is to provide a method, a method, system, and apparatus for generating scene descriptions, proactive question answering, and to accurately describe key content in a scene and accurately answer questions raised by users.
[0004] To address the aforementioned technical problems, this application provides a method for generating scene descriptions, comprising: performing target recognition on video frames in a video to obtain all targets in the video frames; performing windowing processing on the video to obtain multiple video windows, and constructing an anchor point map based on the targets in all the video frames within the video windows; obtaining the selection probability of the targets in the anchor point map of the current video window in a predetermined order, selecting targets whose selection probability exceeds a selection threshold as targets to be described, and generating a relevant description of the targets to be described.
[0005] To address the aforementioned technical problems, another technical solution adopted in this application is to provide an active question-and-answer method, comprising: obtaining relevant descriptions of a video using the method for generating scene descriptions in the above technical solution; and outputting answers based on the relevant descriptions and user questions.
[0006] To address the aforementioned technical problems, another technical solution adopted in this application is: providing a scene description generation system, comprising: a target recognition module, used to perform target recognition on each video frame in a video to obtain all targets in each video frame; a processing module, used to perform windowing processing on the video to obtain multiple video windows, and construct an anchor point map based on the targets in the video frames of each video window; a selection module, used to obtain the selection probability of each target in the anchor point map of the current video window in a predetermined order, and select targets whose selection probability exceeds a threshold as targets to be described; and a description module, used to generate relevant descriptions of the targets to be described.
[0007] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide an electronic device, including a memory and a processor coupled to each other, wherein the memory stores program instructions, and the processor is used to execute the program instructions to implement the method of generating scene description or the active question answering method in the above-mentioned technical solution.
[0008] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a storage device storing program instructions that can be executed by a processor, wherein the program instructions are used to implement the method for generating scene descriptions or the active question-and-answer method in the above-mentioned technical solution.
[0009] The beneficial effects of this application are as follows: Unlike existing technologies, the method for generating scene descriptions proposed in this application obtains multiple video windows by performing windowing processing on the video, and constructs an anchor point map based on the targets in all video frames of each video window. The method then calculates and filters out the targets to be described in each anchor point map and generates relevant descriptions. This method avoids directly performing descriptions on all targets, ensuring the accuracy of the scene description while avoiding the generation of a large amount of redundant description content, thus improving the efficiency of scene description. Attached Figure Description
[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:
[0011] Figure 1 This is a flowchart illustrating one implementation method of generating a scene description according to this application;
[0012] Figure 2 anchor point Figure 1 A schematic diagram of the implementation method;
[0013] Figure 3 This is a flowchart illustrating step S103 corresponding to one implementation method;
[0014] Figure 4 This is a schematic diagram of one embodiment of the scenario description generation system of this application;
[0015] Figure 5 This is a schematic diagram of the structure of one embodiment of the electronic device of this application;
[0016] Figure 6 This is a schematic diagram of one embodiment of the storage device of this application. Detailed Implementation
[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0018] Please see Figure 1 , Figure 1 This is a flowchart illustrating one embodiment of the method for generating scene descriptions according to this application. The method includes:
[0019] S101: Perform target recognition on video frames in the video to obtain all targets in the video frames.
[0020] Specifically, the implementation process of step S101 above includes: acquiring a video from which a scene description needs to be generated, the video consisting of multiple consecutive video frames, and performing target recognition on each video frame to obtain all targets in each video frame, including objects and text in each video frame. Here, a video frame is equivalent to an image. Specifically, each video frame in the above video can be input into a convolutional neural network to identify objects and text in each video frame. Using convolutional neural networks to process images to identify objects and text in images is a common image processing method, and the specific implementation process will not be elaborated here. Optionally, this step can also first construct a target recognition model, and then input each video frame in the video into the target recognition model to identify objects and text in each video frame. The step of constructing the target recognition model includes: acquiring multiple training images, which include manually labeled objects and text; further, using the above multiple training images to train and obtain the target recognition model, the specific process of which will not be elaborated here.
[0021] S102: Perform windowing processing on the video to obtain multiple video windows, and construct an anchor point map based on the targets in all video frames of the video windows.
[0022] Specifically, the implementation process of step S102 above includes: windowing the video to obtain multiple video windows. Specifically, based on the fact that the video includes multiple video frames, the video is divided into multiple video windows, each video window being considered a video segment. Each video window contains video frames with the same number of frames, and the video frames contained in adjacent video windows do not overlap, thus avoiding multiple processing of the same video frame and reducing computational costs. By dividing the video into multiple video windows containing the same number of video frames, processing each video frame individually can be avoided, improving processing efficiency and saving costs while ensuring the accuracy of video processing. Furthermore, the number of video frames contained in each video window can be determined according to the needs of the actual application, and is not limited here. Further, an anchor point map is constructed based on the targets in all video frames within each video window. Specifically, in the anchor point map, each target is connected to its nearest target; between video frames and their adjacent video frames, the same target is connected in the time dimension. Please refer to [link to relevant documentation]. Figure 2 , Figure 2 anchor point Figure 1 A schematic diagram of the implementation method, Figure 2 The anchor point graph shown is constructed based on a video window containing k video frames. The first target f in the first video frame of this video window... 1，1 With the second target f in this video frame 1，2 The closest target is f. 1，1 With the second target f 1，2 Connected. The first target f in the first video frame. 1，1 The target f in the second adjacent video frame 2，1 They are connected in the time dimension. The connection relationships between other objectives can be referred to the connection relationships between the objectives mentioned above, and will not be repeated here.
[0023] S103: Obtain the selection probability of the target in the anchor point map of the current video window in a predetermined order, take the target whose selection probability exceeds the selection threshold as the target to be described, and generate the relevant description of the target to be described.
[0024] For details, please refer to Figure 3 , Figure 3 This is a flowchart illustrating step S103 in one embodiment. Step S103 specifically includes:
[0025] S201: Obtain the selection probability of each target in the anchor point graph of the current video window in a predetermined order.
[0026] In one embodiment, the specific implementation steps of step S201 include: obtaining the initial probability of each target in the anchor point graph of the current video window. Specifically, when the current video window is the first video window among all video windows, the initial probability of each target in the current video window is used as the selection probability to facilitate the execution of step S202. When the current video window is any video window after the first video window, after obtaining the initial probability of each target in the anchor point graph of the current video window, the method further includes: for each target, in response to the fact that the current target has been selected as the target to be described in the video windows preceding the current video window, obtaining the selection probability based on the initial probability of the current target in the current video window and a first preset value. The selection probability is related to the difference between the initial probability and the first preset value. Specifically, for target A in the current video window, if target A has been selected as the target to be described in the video windows preceding the current video window, the initial probability of target A in the current video window is subtracted from the first preset value to obtain the selection probability of target A in the current video window. If target A is selected as the target to be described in multiple video windows preceding the current video window, the initial probability of target A in the current video window is subtracted by a corresponding number of first preset values. For example, if target A is selected as the target to be described in two video windows preceding the current video window, the initial probability of target A in the current video window is subtracted by two first preset values to obtain the selection probability of target A in the current video window. It should be noted that if the selection probability of the target in the current video window is less than 0, then 0 is taken as the selection probability of the target in the current video window. Optionally, in this embodiment, the first preset value is 10%. Of course, in other embodiments, the first preset value can also be changed according to specific circumstances. Obtaining the selection probability based on the initial probability of the current target in the current video window and the first preset value helps to avoid selecting the same target as the target to be described in multiple video windows, thereby preventing the generation of a large number of redundant descriptions.
[0027] In another embodiment, the step of obtaining the initial probability of each target in the anchor graph of the current video window includes: inputting the anchor graph of the current video window into a judgment model to obtain the initial probability of each target. Prior to obtaining the initial probability of each target in the anchor graph of the current video window, the method includes: acquiring multiple training videos. Each training video includes manually labeled tags, which include video frames corresponding to key events in the training video, targets, and event descriptions. Further, the judgment model is trained using multiple training videos. The tag generation process includes: after watching the entire training video, the labeler labels the video frames corresponding to key events and targets in the training video, and provides event descriptions based on the video frames corresponding to the key events and targets. The video frames corresponding to the key events can be any video frame in the training video, or a video segment composed of multiple video frames. By constructing a judgment model to facilitate the processing of the anchor graph, the initial probability of each target in the anchor graph is obtained, and the selection probability of each target in the anchor graph can be obtained based on this initial probability.
[0028] S202: Select targets whose selection probability exceeds the selection threshold as targets to be described.
[0029] Specifically, step S202 includes comparing the selection probability obtained in step S201 with a selection threshold, and designating targets with selection probabilities exceeding the selection threshold as targets to be described, thus facilitating the execution of step S203. The selection threshold can be 50%, meaning that when the selection probability of a target in the anchor point graph exceeds 50%, that target is designated as the target to be described. Of course, in other embodiments, the selection threshold can also be other values greater than 0, and this is not limited.
[0030] S203: Generate a relevant description of the target to be described.
[0031] Specifically, before step S203, the process includes: acquiring multiple training videos, each containing manually labeled tags, which include video frames corresponding to key events in the training video, the target, and event descriptions. Further, a description model is trained using these multiple training videos.
[0032] Further, the specific implementation process of step S203 includes: obtaining the selected video frame where the target to be described is located from the current video window. Based on the target to be described in the selected video frame, all other targets besides the target to be described in the selected video frame, and the target to be described in the remaining video frames in the current video window, a relevant description of the target to be described is obtained. Specifically, the anchor point map of the current video window is input into the description model to obtain the hidden state features of the target to be described in the selected video frame, all other targets besides the target to be described in the selected video frame, and the target to be described in the remaining video frames in the current video window, based on the anchor point map. The fully connected neural network (FC) in the description model then generates a relevant description centered on the target to be described based on these hidden state features. Combining the different features of other targets in the same video frame and the target to be described in different video frames to generate relevant descriptions helps to obtain the relationship between the target to be described and other targets, as well as the change features of the target to be described within the video window, thus ensuring the accuracy of the generated relevant descriptions.
[0033] Furthermore, the specific implementation process of step S203 also includes: generating a confidence score for the relevant description. In response to a confidence score exceeding a confidence threshold, the relevant description is output. In response to a confidence score lower than or equal to the confidence threshold, the relevant description is stored in a database, and the description stored in the database, along with its corresponding video window, can be used as training videos to train the judgment model in step S201. The confidence threshold can be 50%. Of course, in other embodiments, the confidence threshold can also be other values greater than 0, and this is not limited. Outputting descriptions with confidence scores higher than the confidence threshold helps avoid outputting a large number of redundant descriptions and improves the efficiency of generating scene descriptions.
[0034] The proposed method for generating scene descriptions involves windowing a video to obtain multiple video windows, constructing an anchor point map based on targets in all video frames within each window, and then calculating and filtering out the targets to be described in each anchor point map to generate relevant descriptions. This method avoids directly applying descriptions to all targets, ensuring accuracy in scene description while preventing the generation of a large amount of redundant description content, thus improving the efficiency of scene description.
[0035] In another embodiment, Figure 3 The step S201, which involves obtaining the selection probability of each target in the anchor point graph of the current video window, includes: obtaining the initial probability of each target in the anchor point graph of the current video window. This process can be referred to... Figure 3Step S201 will not be elaborated here. Further, for each target, in response to the fact that the video windows preceding the current video window have obtained relevant descriptions of the current target, and the confidence score of the relevant descriptions is lower than the confidence threshold, a selection probability is obtained based on the initial probability and a second preset value of the current target in the current video window. This process helps to avoid omitting targets that need to be described, preventing inaccurate descriptions. The selection probability is related to the sum of the initial probability and the second preset value. Specifically, in the video windows preceding the current video window, if the selection probability of the current target exceeds the selection threshold and it is used as the target to be described, the description model generates a relevant description centered on that target and a corresponding confidence score. If the confidence score is lower than the confidence threshold, then the selection probability of that target in the current video window is the sum of the initial probability and the second preset value. Optionally, in this embodiment, the second preset value can be 10%. Of course, in other embodiments, the second preset value can also be changed according to the actual situation.
[0036] In addition, in this embodiment, before the step of obtaining the selection probability based on the initial probability of the current target in the current video window and the second preset value, the method includes: determining whether a relevant description of the current target has been obtained in a video window preceding the current video window, and whether the number of times the confidence score of the relevant description is lower than the confidence threshold exceeds a threshold. If yes, that is, if a relevant description of the current target has been obtained in a video window preceding the current video window, and the number of times the confidence score of the relevant description is lower than the confidence threshold exceeds the threshold, then the selection probability of the target is set to be lower than the selection threshold, for example, 0 or other values lower than the selection threshold. If no, then the selection probability is obtained based on the initial probability of the current target in the current video window and the second preset value. The threshold can be a natural number such as 2, 3, or 4; preferably, when the second preset value is 10%, the threshold can be set to 3. For ease of understanding, the threshold for the number of attempts is set to 3. The first video window, the second video window, and the third video window are three video windows in the video, with the first video window preceding the second video window and the second video window preceding the third video window. If the current target obtains relevant descriptions in the first, second, and third video windows, but the confidence scores of the relevant descriptions are all lower than the confidence threshold, then the selection probability of the current target in the video window after the third video window is set to be lower than the selection threshold.
[0037] It should be noted that if a relevant description of the current target has been obtained in the current video window, and the confidence score of the relevant description is lower than the confidence threshold, the larger the number of times threshold is, the greater the probability of the current target being selected in the video window after the current video window. Therefore, in practical applications, it is necessary to avoid the number of times threshold being too large, which may affect the selection probability of the target.
[0038] This invention also proposes an active question-answering method, which includes obtaining relevant descriptions of a video using the scene description generation method described in the above embodiments. Further, an answer is output based on the relevant descriptions and the user's question. Specifically, this method can filter relevant scene descriptions based on the user's question and input the anchor graphs corresponding to these relevant scene descriptions into a visual question-answering model for reasoning, thereby quickly answering the user's question. Additionally, when no scene description relevant to the user's question can be matched, the visual question-answering model processes all anchor graphs to provide an answer to the relevant question. It should be noted that when there is a difference between the keywords in the user's question and the name of the target in the extracted video frame, the keywords in the user's question can be matched with the name of the target in the video frame to obtain a similarity score. This similarity score is compared with a similarity threshold. If the similarity score is greater than the similarity threshold, it is considered that the keywords in the user's question match the name of the target in the video frame, and the user's question is answered based on the scene description related to that target.
[0039] Please see Figure 4 , Figure 4 This is a schematic diagram of one embodiment of the scene description generation system of this application. The scene description generation system includes: a target recognition module 11, a processing module 12, a selection module 13, and a description module 14. The target recognition module 11 is connected to the processing module 12 and is used to perform target recognition on each video frame in the video to obtain all targets in each video frame, and input the video with target recognition to the processing module. The processing module 12 is connected to the selection module 13 and is used to perform windowing processing on the video to obtain multiple video windows, and construct an anchor point map based on the targets in the video frames of each video window. This anchor point map is input to the selection module 13. The video frames contained in two adjacent video windows do not overlap. The selection module 13 is connected to the description module 14 and is used to obtain the selection probability of each target in the anchor point map of the current video window in a predetermined order, and input targets with selection probabilities exceeding a selection threshold as targets to be described into the description module 14. The description module 14 is used to generate relevant descriptions of the targets to be described.
[0040] Specifically, the step of selecting module 13 to obtain the selection probability of each target in the anchor point graph of the current video window includes: obtaining the initial probability of each target in the anchor point graph of the current video window; for each target, in response to the fact that the current target has been selected as the target to be described in the video window before the current video window, obtaining the selection probability based on the initial probability of the current target in the current video window and a first preset value; wherein, the selection probability is related to the difference between the initial probability and the first preset value.
[0041] The step of selecting module 13 to obtain the initial probability of each target in the anchor point graph of the current video window includes: inputting the anchor point graph of the current video window into the judgment model to obtain the initial probability of each target. Prior to obtaining the initial probability of each target in the anchor point graph of the current video window, the process includes: acquiring multiple training videos; each training video includes manually labeled tags, which include the video frame corresponding to the key event in the training video, the target, and the event description; and training the judgment model using the multiple training videos.
[0042] Specifically, the step of selecting module 13 to obtain the selection probability of each target in the anchor point graph of the current video window further includes: obtaining the initial probability of each target in the anchor point graph of the current video window. For each target, in response to the fact that the video windows preceding the current video window have obtained relevant descriptions of the current target, and the confidence score of the relevant descriptions is lower than the confidence threshold, the selection probability is obtained based on the initial probability of the current target in the current video window and a second preset value; wherein, the selection probability is related to the sum of the initial probability and the second preset value.
[0043] Specifically, before the step of obtaining the selection probability based on the initial probability and the second preset value of the current target in the current video window, the selection module 13 includes: determining whether the video window before the current video window has obtained a relevant description of the current target, and whether the number of times the confidence score of the relevant description is lower than the confidence threshold exceeds the number threshold; if so, the selection probability of the target is set to be lower than the selection threshold; otherwise, the selection probability is obtained based on the initial probability and the second preset value of the current target in the current video window.
[0044] Specifically, the description module 14 generates a confidence score for the relevant description. In response to a confidence score exceeding a confidence threshold, the relevant description is output.
[0045] Specifically, the step of the description module 14 in generating a relevant description of the target to be described includes: obtaining the selected video frame in which the target to be described is located from the current video window; and obtaining a relevant description of the target to be described based on the target to be described in the selected video frame, all other targets besides the target to be described in the selected video frame, and the targets to be described in the other video frames in the current video window.
[0046] Please see Figure 5 , Figure 5This is a schematic diagram of the structure of an embodiment of the electronic device of this application. The electronic device includes a memory 20 and a processor 30 coupled to each other. The memory 20 stores program instructions, and the processor 30 is used to execute the program instructions to implement the steps of the method for generating scene descriptions or the active question-and-answer method described in any of the above embodiments. Specifically, the electronic device includes, but is not limited to, desktop computers, laptops, tablets, servers, etc., and is not limited thereto. In addition, the processor 30 may also be called a CPU (Center Processing Unit). The processor 30 may be an integrated circuit chip with signal processing capabilities. The processor 30 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor may be a microprocessor or any conventional processor. In addition, the processor 30 may be implemented by integrated circuit chips.
[0047] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of a storage device according to an embodiment of the present application. The storage device 50 stores program instructions 60 that can be executed by a processor. The program instructions 60 are used to implement the steps of the method for generating scene descriptions or the active question-and-answer method described in any of the above embodiments.
[0048] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A method for generating scene descriptions, characterized in that, include: Target recognition is performed on video frames in the video to obtain all targets in the video frames; The video is windowed to obtain multiple video windows, and an anchor point map is constructed based on the target in all the video frames in the video windows; wherein, in the anchor point map, each target is connected to its nearest target, and the same target is connected in the time dimension between video frames and their adjacent video frames; The selection probability of the target in the anchor point graph of the current video window is obtained in a predetermined order. Targets whose selection probability exceeds the selection threshold are taken as targets to be described, and relevant descriptions of the targets to be described are generated.
2. The method according to claim 1, characterized in that, The step of obtaining the selection probability of the target in the anchor point map of the current video window includes: Obtain the initial probability of each target in the anchor point graph of the current video window; For each target, in response to the fact that the current target has been selected as the target to be described in the video window preceding the current video window, the selection probability is obtained based on the initial probability of the current target in the current video window and a first preset value; wherein, the selection probability is related to the difference between the initial probability and the first preset value.
3. The method according to claim 1, characterized in that, Also includes: Generate a confidence score for the relevant description; In response to the confidence score exceeding the confidence threshold, the relevant description is output.
4. The method according to claim 3, characterized in that, The step of obtaining the selection probability of the target in the anchor point map of the current video window includes: Obtain the initial probability of each target in the anchor point graph of the current video window; For each target, in response to the fact that the video window preceding the current video window has obtained a relevant description of the current target, and the confidence score of the relevant description is lower than the confidence threshold, the selection probability is obtained based on the initial probability and the second preset value of the current target in the current video window; wherein, the selection probability is related to the sum of the initial probability and the second preset value.
5. The method according to claim 4, characterized in that, Before the step of obtaining the selection probability based on the initial probability and the second preset value of the target in the current video window, the following steps are included: Determine whether the video windows preceding the current video window have obtained a relevant description of the current target, and whether the number of times the confidence score of the relevant description is lower than the confidence threshold exceeds the number threshold; If so, the selection probability of the target is set to be lower than the selection threshold; Otherwise, the selection probability is obtained based on the initial probability and the second preset value of the target in the current video window.
6. The method according to claim 1, characterized in that, The step of generating a relevant description of the target to be described includes: Obtain the selected video frame containing the target to be described from the current video window; Based on the target to be described in the selected video frame, all other targets in the selected video frame besides the target to be described, and the target to be described in the remaining video frames in the current video window, the relevant description of the target to be described is obtained.
7. The method according to claim 2, characterized in that, The step of obtaining the initial probability of each target in the anchor point graph of the current video window includes: inputting the anchor point graph of the current video window into the judgment model to obtain the initial probability of each target; Before the step of obtaining the initial probability of each target in the anchor point graph of the current video window, the method includes: Multiple training videos are acquired; each training video includes manually labeled tags, the tags including video frames, targets, and event descriptions corresponding to key events in the training video; The judgment model is trained using the multiple training videos.
8. The method according to claim 1, characterized in that, The video frames contained in two adjacent video windows do not overlap.
9. A method for proactive question answering, characterized in that, include: A relevant description of a video is obtained using the method for generating scene descriptions as described in any one of claims 1-8; The answer is output based on the relevant description and the user's question.
10. A scene description generation system, characterized in that, include: The target recognition module is used to perform target recognition on each video frame in the video to obtain all targets in each video frame; The processing module is used to perform windowing processing on the video to obtain multiple video windows, and to construct an anchor point map based on the target in the video frame in each video window; wherein, in the anchor point map, each target is connected to its nearest target, and the same target is connected in the time dimension between the video frame and its adjacent video frames. The selection module is used to obtain the selection probability of each target in the anchor point map of the current video window in a predetermined order, and to select targets whose selection probability exceeds the selection threshold as targets to be described. The description module is used to generate relevant descriptions of the target to be described.
11. An electronic device, characterized in that, It includes a memory and a processor coupled to each other, the processor storing program instructions, and the memory executing the program instructions to implement the method for generating scene descriptions as described in any one of claims 1-8 or the method for active question answering as described in claim 9.
12. A storage device, characterized in that, The system stores program instructions that can be executed by a processor, the program instructions being used to implement the method for generating scene descriptions as described in any one of claims 1-8 or the method for proactive question answering as described in claim 9.