A video target detection model training method, device, equipment and medium

Abstract

The embodiment of the application provides a video target detection model training method and device, equipment and medium. In the embodiment of the application, in addition to training the original video target detection model based on the sample images with labeled labels in the image sample set, the original video target detection model is also trained based on the target reference frames without labeled labels in the video sample set and with a confidence greater than a preset first threshold. That is, the application adopts a semi-supervised learning method, that is, even if all video frames are not labeled, a high-precision video target detection model can be trained, and the accuracy of target detection is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device and medium for training a video object detection model. Background Technology

[0002] With the continuous development of target tracking and detection technology, video target detection technology, as an important branch of target tracking and detection technology, has been widely used in practical scenarios such as intelligent video surveillance, autonomous driving, human-machine interfaces, and video artificial intelligence analysis. Specifically, this video target detection technology is used to locate the category of objects appearing in each target video frame of a video sequence and to determine the object's position within the target video frame.

[0003] In related technologies, pre-trained video object detection models are generally used to identify the categories and locations of objects appearing in video frames. Training these models requires labeling the video frames used for training. However, labeling the location and category of all objects in a single video frame takes approximately 10 seconds, while a 10-second video typically consists of 400-600 frames. Therefore, the labeling time is too long, and the large number of frames to label makes labeling difficult. If the video object detection model is trained based on a small number of labeled video frames, the model's accuracy will be low, resulting in low precision in identifying object categories and locations. Summary of the Invention

[0004] This application provides a video object detection model training method, apparatus, device, and medium to solve the problems in the prior art where the workload of annotating video frames is large and training the video object detection model based on a small number of annotated video frames results in low accuracy of the video object detection model and low accuracy in object recognition.

[0005] This application provides a method for training a video object detection model, the method comprising:

[0006] For any sample image in a pre-saved image sample set, the sample image is input into the original video object detection model to obtain the first position information of each object in the sample image and the first confidence level of the category to which each object belongs;

[0007] For any video frame in the pre-saved video sample set, the video frame is input into the original video target detection model to obtain the second position information of each object in the video frame and the second confidence score of the category to which each object belongs; if the second confidence score of any object is greater than a preset first threshold, the video frame is used as the target reference frame.

[0008] The target reference frame is input into the original video target detection model to obtain the third position information of each object in the target reference frame and the third confidence score of the category to which each object belongs;

[0009] A first loss value is determined based on the target location information of each object identified in the label of each sample image and the target confidence of the target category of each object, and the first location information and the first confidence of each object. A second loss value is determined based on the second location information and the second confidence of each object in the target reference frame, and the third location information and the third confidence of each object.

[0010] The parameters of the original video target detection model are updated based on the first loss value and the second loss value.

[0011] Furthermore, after using the video frame as the target reference frame, the method further includes:

[0012] Based on the second confidence level of each object in other video frames adjacent to the video frame, and whether the other video frames are content-continuous with the video frame, it is determined whether to identify the other video frames as target reference frames.

[0013] Further, determining whether to identify the other video frames as target reference frames based on the second confidence level corresponding to each object in other video frames adjacent to the video frame, and whether the other video frames are content-continuous with the video frame, includes:

[0014] For any other video frame preceding or following the video frame adjacent to the video frame, if the other video frame is not the target reference frame, the second confidence level corresponding to any object in the other video frame is less than the preset first threshold and greater than the preset second threshold, and the content of the other video frame is continuous with that of the video frame, then the other video frame is determined as the target reference frame.

[0015] Furthermore, the method also includes:

[0016] The other video frames are updated to the video frame, and the previous or next other video frame adjacent to the video frame is determined as the target reference frame until the other video frame is the target reference frame, or the second confidence of each object in the other video frame is less than the preset second threshold, or the other video frame is not continuous with the content of the video frame.

[0017] Further, determining that the other video frames are continuous with the content of the video frames includes:

[0018] Determine whether there is a target cross-union ratio (IOU) greater than a preset IOU threshold among the IOUs of each object in the other video frames and each object in the target reference frame;

[0019] If so, then the other video frames are continuous with the content of the video frame.

[0020] Further, updating the parameters of the original video target detection model based on the first loss value and the second loss value includes:

[0021] The parameters of the original video target detection model are updated based on the sum of the products of the first loss value, the second loss value, and their corresponding weight values.

[0022] This application provides a video object detection model training device, the device comprising:

[0023] The acquisition module is used to: input any sample image from a pre-saved image sample set into the original video object detection model to acquire the first position information of each object in the sample image and the first confidence score of the category to which each object belongs; input any video frame from a pre-saved video sample set into the original video object detection model to acquire the second position information of each object in the video frame and the second confidence score of the category to which each object belongs; if the second confidence score of any object is greater than a preset first threshold, the video frame is used as a target reference frame; input the target reference frame into the original video object detection model to acquire the third position information of each object in the target reference frame and the third confidence score of the category to which each object belongs;

[0024] The determination module is used to determine a first loss value based on the target location information of each object identified in the label of each sample image and the target confidence of the target category of each object, and the first location information and the first confidence of each object; and to determine a second loss value based on the second location information and the second confidence of each object in the target reference frame, and the third location information and the third confidence of each object.

[0025] An update module is used to update the parameters of the original video target detection model based on the first loss value and the second loss value.

[0026] Furthermore, the acquisition module is also used to determine whether to identify the other video frames as target reference frames based on the second confidence level corresponding to each object in other video frames adjacent to the video frame, and whether the other video frames are content-continuous with the video frame.

[0027] Furthermore, the acquisition module is specifically used to determine the other video frame as the target reference frame if the other video frame adjacent to the video frame is not the target reference frame, the second confidence level corresponding to any object in the other video frame is less than the preset first threshold and greater than the preset second threshold, and the other video frame is continuous with the content of the video frame.

[0028] Furthermore, the acquisition module is also used to update the other video frames to the video frame, and continue to determine whether the previous or next other video frame adjacent to the video frame is the target reference frame, until the other video frame is the target reference frame, or the second confidence of each object in the other video frame is less than a preset second threshold, or the other video frame is not continuous with the content of the video frame.

[0029] Furthermore, the acquisition module is specifically used to determine whether there is a target cross-union ratio (IOU) greater than a preset IOU threshold among the IOUs of each object in the other video frames and each object in the target reference frame; if so, the other video frames and the video frame content are continuous.

[0030] Furthermore, the update module is specifically used to update the parameters of the original video target detection model based on the sum of the products of the first loss value, the second loss value, and their respective weight values.

[0031] This application provides an electronic device including a processor, which executes a computer program stored in a memory to implement the steps of the video object detection model training method as described above.

[0032] This application provides a computer-readable storage medium storing a computer program executable by a terminal, which, when run on the terminal, causes the terminal to perform the steps of any of the above-described video object detection model training methods.

[0033] In this embodiment, for any sample image in a pre-saved image sample set, the sample image is input into the original video object detection model to obtain the first position information of each object in the sample image and the first confidence score of the category to which each object belongs. For any video frame in a pre-saved video sample set, the video frame is input into the original video object detection model to obtain the second position information of each object in the video frame and the second confidence score of the category to which each object belongs. If the second confidence score of any object is greater than a preset first threshold, the video frame is used as a target reference frame, and the target reference frame is input into the original video object detection model. In this model, the third position information of each object in the target reference frame and the third confidence level of the category to which each object belongs are obtained. Based on the target position information and target confidence level of each object's target category identified in the label of each sample image, and the corresponding first position information and first confidence level of each object, a first loss value is determined. Based on the second position information and second confidence level of each object in the target reference frame, and the corresponding third position information and third confidence level of each object, a second loss value is determined. Based on the first and second loss values, the parameters of the original video object detection model are updated. In this embodiment, in addition to training the original video object detection model based on labeled sample images from the image sample set, the model is also trained based on unlabeled target reference frames in the video sample set with confidence levels greater than a preset first threshold. That is, this application uses a semi-supervised learning method, which can train a high-precision video object detection model even without labeling all video frames, thus improving the accuracy of object detection. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0035] Figure 1 This is a schematic diagram illustrating the process of a video object detection model training method provided in an embodiment of this application;

[0036] Figure 2 A schematic diagram illustrating the process of determining a target reference frame provided in some embodiments of this application;

[0037] Figure 3 A schematic diagram illustrating the display of a video frame provided for some embodiments of this application;

[0038] Figure 4A schematic diagram illustrating the training process of a video object detection model provided in some embodiments of this application;

[0039] Figure 5 This application provides a schematic diagram of a target detection process for some embodiments;

[0040] Figure 6 A schematic diagram illustrating the effect of model training results provided in some embodiments of this application;

[0041] Figure 7 A schematic diagram of a video object detection model training device provided in some embodiments of this application;

[0042] Figure 8 This is a schematic diagram of the structure of an electronic device provided in some embodiments of this application. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail 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 in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0044] In this embodiment, for any sample image in a pre-saved image sample set, the sample image is input into the original video object detection model to obtain the first position information of each object in the sample image and the first confidence score of the category to which each object belongs. For any video frame in a pre-saved video sample set, the video frame is input into the original video object detection model to obtain the second position information of each object in the video frame and the second confidence score of the category to which each object belongs. If the second confidence score of any object is greater than a preset first threshold, the video frame is used as a target reference frame, and the target reference frame is input into the original video object detection model. In this model, the third position information of each object in the target reference frame and the third confidence score of the category to which each object belongs are obtained. Based on the target position information and target confidence score of each object's target category identified in the label of each sample image, and the corresponding first position information and first confidence score of each object, a first loss value is determined. Based on the second position information and second confidence score of each object in the target reference frame, and the corresponding third position information and third confidence score of each object, a second loss value is determined. Based on the first loss value and the second loss value, the parameters of the original video object detection model are updated.

[0045] In order to train a high-precision video object detection model through semi-supervised learning and improve the accuracy of object detection, this application provides a video object detection model training method, apparatus, device and medium.

[0046] Example 1:

[0047] Figure 1 This application provides a schematic diagram of a video object detection model training method, which includes:

[0048] S101: For any sample image in the pre-saved image sample set, input the sample image into the original video object detection model to obtain the first position information of each object in the sample image and the first confidence level of the category to which each object belongs.

[0049] The video target detection model training method provided in this application is applied to electronic devices, such as PCs and servers.

[0050] In this embodiment of the application, in order to obtain a video target detection model with high recognition accuracy, an image sample set is pre-stored, and the original video target detection model is trained based on each sample image contained in the pre-stored image sample set.

[0051] To facilitate subsequent training of the video object detection model, each sample image in the image sample set is labeled. These labels identify the target location information of each object in the corresponding sample image and the target confidence score for each object's target category. For each object, the target location information is its actual location in the sample image, and the target category is its actual category. Since the labeling of the sample images uses the actual category of the object, and for any given object, its actual category can only be one, the target confidence score for that object's target category is 1.

[0052] In this embodiment of the application, in order to train the original video object detection model, for any sample image in the pre-saved image sample set, the sample image is input into the original video object detection model to obtain the first position information of each object in the sample image and the first confidence score of the category to which each object belongs, as output by the original video object detection model. For each object in the sample image, the first position information of the object in the sample image is the position information of the object in the sample image predicted by the original video object detection model, and the first confidence score of the category to which the object belongs in the sample image is the probability value of the object belonging to the category with the highest probability predicted by the original video object detection model. For example, if a sample image containing object A and object B is input into the original video object detection model, and the original video object detection model determines that object A belongs to the apple category with a confidence score of 0.9, object A belongs to the pear category with a confidence score of 0.7, object A belongs to the ping-pong ball category with a confidence score of 0.2, etc., then the original video object detection model outputs that the first confidence score of object A belonging to the apple category in the sample image is 0.9; if the original video object detection model determines that object B belongs to the cup category with a confidence score of 0.85, object B belongs to the teapot category with a confidence score of 0.5, then the original video object detection model outputs that the first confidence score of object B belonging to the cup category in the sample image is 0.85.

[0053] S102: For any video frame in the pre-saved video sample set, input the video frame into the original video target detection model to obtain the second position information of each object in the video frame and the second confidence level of the category to which each object belongs; if the second confidence level of any object is greater than the preset first threshold, the video frame is used as the target reference frame.

[0054] In this embodiment, to obtain a video object detection model with high recognition accuracy, a video sample set is pre-stored. This video sample set contains at least one video, and each video contains multiple video frames. When training the video object detection model, for any video frame in the pre-stored video sample set, the video frame is input into the original video object detection model. The second position information of each object in the video frame and the second confidence score of each object's category are obtained from the output of the original video object detection model. Specifically, for each object in the video frame, the second position information of the object in the video frame is the position information of the object in the video frame predicted by the original video object detection model, and the second confidence score of the object's category is the probability value predicted by the original video object detection model that the object belongs to the category with the highest probability. In this embodiment, none of the video frames in the pre-stored video sample set are labeled.

[0055] To facilitate training the original video object detection model based on unlabeled video frames, reliable target reference frames can be determined based on the second position information of each object and the second confidence score of each object's category in each video frame output by the original video object detection model. In this application, reliable output means video frames where the output object's second position information and the object's category are very close to the object's actual position information and actual category.

[0056] In this embodiment of the application, in order to determine a reliable target reference frame, a first threshold is preset in the electronic device. For any video frame in the pre-saved video sample set, after obtaining the second position information of each object in the video frame and the second confidence level of the category to which each object belongs, the second confidence level of the category to which each object belongs in the video frame is compared with the preset first threshold. If the second confidence level of any object is greater than the preset first threshold, it indicates that the output result corresponding to the video frame is reliable, and the video frame is used as the target reference frame.

[0057] After judging each video frame of each video in the video sample set using the above method, each target reference frame can be determined.

[0058] S103: Input the target reference frame into the original video target detection model to obtain the third position information of each object in the target reference frame and the third confidence level of the category to which each object belongs.

[0059] In this embodiment, since the output of the original video object detection model for the target reference frame is reliable—meaning that each object in the target reference frame predicted by the original video object detection model contains target objects whose actual position information and actual class classification are very close to those of the object in the output second position information and second confidence score of the classification—it is possible to use the second position information and second confidence score of each object in the target reference frame as the label for that target reference frame after determining each target reference frame in the video sample set. This target reference frame is then input back into the original video object detection model to obtain the third position information and third confidence score of each object's classification in the target reference frame output by the original video object detection model. Specifically, for each object in the target reference frame, the third position information of the object in the target reference frame is the position information of the object in the target reference frame predicted by the original video object detection model. The third confidence score of the object's classification in the target reference frame is the probability value of the object belonging to the highest probability class predicted by the original video object detection model.

[0060] S104: Determine a first loss value based on the target location information of each object identified in the label of each sample image and the target confidence of the target category of each object, and the first location information and the first confidence of each object corresponding to each object; determine a second loss value based on the second location information and the second confidence of each object in the target reference frame, and the third location information and the third confidence of each object corresponding to each object.

[0061] In this embodiment of the application, in order to complete the training of the original video target detection model, the loss value corresponding to the sample image in the image sample set can be determined based on the target location information and target confidence of each object in the label of each sample image, and the first location information and first confidence of each object in the corresponding sample image. The loss value corresponding to the image sample set is also the first loss value.

[0062] Specifically, when determining the first loss value corresponding to the image sample set, for each object in each sample image in the image sample set, the loss value corresponding to the object in the sample image can be determined based on the difference between the target location information of the object in the label of the sample image and the first location information of the object in the sample image, the difference between the target confidence of the target category of the object in the label of the sample image and the first confidence of the category to which the object belongs, and the sum of the loss values ​​corresponding to each object in the sample image is determined as the loss value corresponding to the sample image, and the first sum of the loss values ​​corresponding to each sample image is determined as the first loss value.

[0063] In this application, the first loss value can be determined using the following formula:

[0064] ;

[0065] in, The first loss value, Let M be the loss value corresponding to the j-th sample image, and M be the total number of sample images in the image sample set. The difference between the target location information of object i identified in the label of the j-th sample image and the first location information of the i-th object in the j-th sample image. Let P be the difference between the target location information of object i identified in the label of the j-th sample image and the first confidence score of the category to which the i-th object in the j-th sample image belongs, and let P be the total number of objects contained in the j-th sample image. These are preset parameters; specifically, these preset parameters can be set according to requirements.

[0066] In this embodiment of the application, in order to complete the training of the original video object detection model, for each object in each target reference frame, the loss value corresponding to the object in the target reference frame can be determined according to the second position information of the object in the target reference frame and the second confidence of the object's category, and the third position information of the object in the target reference frame and the third confidence of the object's category in the target reference frame. The sum of the loss values ​​corresponding to each object in the target reference frame is determined as the loss value corresponding to the target reference frame. And a second loss value is determined according to the loss value corresponding to each target reference frame.

[0067] Specifically, for each object in each target reference frame, the loss value corresponding to the object in the target reference frame can be determined based on the difference between the second position information and the third position information of the object in the target reference frame, and the difference between the second confidence level and the third confidence level of the object's class in the target reference frame. The loss value corresponding to each object in the target reference frame is then determined as the loss value corresponding to the target reference frame, and the second sum of the loss values ​​corresponding to each target reference frame is determined as the second loss value.

[0068] In this application, the second loss value can be determined using the following formula:

[0069] ;

[0070] in, This is the second loss value. Let H be the loss value corresponding to the m-th video frame, and H be the number of all video frames included in the video sample set. If the m-th video frame is the target reference frame, then this ,otherwise, , The difference between the second position information of object k in the m-th video frame and the third position information of object k in the m-th video frame. Let Q be the difference between the second confidence level and the third confidence level of object k belonging to the class in the m-th video frame, and let Q be the total number of objects in the m-th video frame. These are preset parameters; specifically, these preset parameters can be set according to requirements.

[0071] S105: Update the parameters of the original video target detection model based on the first loss value and the second loss value.

[0072] In this embodiment of the application, after determining the first loss value and the second loss value, the original video object detection model is trained based on the first loss value and the second loss value. The smaller the first loss value and the second loss value, the better the training result of the trained video object detection model and the higher the accuracy of object detection.

[0073] In this embodiment, in addition to training the original video target detection model based on the input of labeled sample images from the image sample set, the original video target detection model is also trained based on unlabeled target reference frames from the video sample set with a confidence level greater than a preset first threshold. That is, this application adopts a semi-supervised learning method, which can train a high-precision video target detection model even without labeling all video frames, thereby improving the accuracy of target detection.

[0074] Example 2:

[0075] To increase the number of reliable target reference frames for the output results and facilitate subsequent improvement of the recognition accuracy of the video target detection model, based on the above embodiments, in this embodiment of the application, after using the video frame as the target reference frame, the method further includes:

[0076] Based on the second confidence level of each object in other video frames adjacent to the video frame, and whether the other video frames are content-continuous with the video frame, it is determined whether to identify the other video frames as target reference frames.

[0077] In this embodiment, since there is a certain correlation between two adjacent video frames, and the features of one of the two adjacent video frames can be used to enhance the features of the other video frame, in order to increase the number of reliable target reference frames and improve the accuracy of the video target detection model, for any video frame in the pre-saved video sample set, if the second confidence of the category to which any object belongs in the video frame is greater than a preset first threshold, the video frame can be determined as a target reference frame. Then, it is determined whether there are any reliable video frames in other video frames adjacent to the video frame, and then the reliable video frames in other video frames adjacent to the video frame are determined as target reference frames.

[0078] Specifically, to determine whether other videos adjacent to a given video frame can serve as a target reference frame, the second confidence level corresponding to each object in those adjacent video frames, and whether those other video frames are content-continuous with the given video frame, can be used to determine whether to designate those other video frames as target reference frames. The higher the second confidence level corresponding to any object in those other video frames, the greater the likelihood that those other video frames will be designated as target reference frames. Furthermore, the greater the likelihood that those other video frames are content-continuous with the given video frame, the greater the likelihood that they will be designated as target reference frames.

[0079] To accurately determine whether other videos adjacent to the video frame can be identified as target reference frames, based on the above embodiments, in this embodiment, determining whether to identify the other video frames as target reference frames according to the second confidence level corresponding to each object in the other video frames adjacent to the video frame, and whether the other video frames are content-continuous with the video frame, includes:

[0080] For any other video frame preceding or following the video frame adjacent to the video frame, if the other video frame is not the target reference frame, the second confidence level corresponding to any object in the other video frame is less than the preset first threshold and greater than the preset second threshold, and the content of the other video frame is continuous with that of the video frame, then the other video frame is determined as the target reference frame.

[0081] In this application, since the second confidence threshold of other videos adjacent to the video frame may be greater than the preset first threshold, meaning that other video frames adjacent to the video frame may have already been identified as target reference frames, it is not necessary to determine whether the other video frame is a target reference frame based on the second confidence level of the other video frame and whether the other video frame is content-continuous with the video frame. The output result of the other video frame is reliable. Therefore, in this embodiment, for the preceding or following other video frame adjacent to the video frame, if the other video frame is a target reference frame, it is not necessary to determine whether the other video frame is a target reference frame based on the second confidence level of each object in the other video frame and whether the other video frame is content-continuous with the video frame. The preceding other video frame adjacent to the video frame is a video frame whose acquisition time is adjacent to and earlier than that of the video frame, and the following other video frame adjacent to the video frame is a video frame whose acquisition time is adjacent to and later than that of the video frame.

[0082] For example, if video frame i is the target reference frame, and the preceding video frame i-1 adjacent to video frame i is also the target reference frame, then it is not necessary to make a judgment on the preceding video frame i-1. If the following video frame i+1 adjacent to video frame i is not the target reference frame, then for the following video frame i+1, it can be determined whether the following video frame i+1 is the target reference frame based on the second confidence level corresponding to each object in the following video frame i+1 and whether the following video frame i+1 and video frame i are content-continuous.

[0083] In this embodiment of the application, for the preceding or following video frame adjacent to the video frame, if the other video frame is not the target reference frame, the operation of determining whether to identify the other video frame as the target reference frame is performed based on the second confidence level corresponding to each object in the other video frames adjacent to the video frame, and whether the other video frames are continuous with the video frame.

[0084] Specifically, a second threshold is preset in the electronic device, wherein the preset second threshold is less than a preset first threshold. In the embodiments of this application, the preset first threshold can be 0.9, 0.8, 0.7, etc., and the preset second threshold can be 0.4, 0.3, etc., values ​​less than the preset first threshold. Specifically, the specific values ​​of the preset first threshold and the preset second threshold can be set according to requirements.

[0085] To determine whether there is a reliable target video frame among other video frames adjacent to the target reference frame, it can be determined whether the second confidence level corresponding to each object in any other video frame adjacent to the target reference frame is less than the preset first threshold and greater than the preset second threshold, and whether the other video frame and the target reference frame are content-continuous. If both are true, it indicates that the output result of the other video frame is reliable, and the other video frame is determined as the target reference frame.

[0086] In this application, it can also be determined whether the second confidence level corresponding to any object in any other video frame adjacent to the target reference frame is less than the preset first threshold and greater than the preset second threshold, and whether the other video frame and the target reference frame are content continuous. If both are true, it indicates that the output result of the other video frame is reliable, and the other video frame is determined as the target reference frame.

[0087] Figure 2 This application provides a schematic diagram illustrating a process for determining a target reference frame in some embodiments. Figure 3 This is a schematic diagram of a video frame provided in some embodiments of this application, and is now specifically for... Figure 2 as well as Figure 3 Please provide an explanation.

[0088] Each video frame from the video sample set is input into the original video object detection model to obtain the second confidence score of the category to which each object belongs in each video frame. For each video frame, if the second confidence score of the category to which any object belongs in the video frame is greater than a preset first threshold, then the video frame is determined as a target reference frame. In this embodiment, the target reference frame can be stored in a pseudo-label set and a reliable label set, respectively. Figure 3 As shown, the target reference frames whose second confidence score of the category to which any object belongs in the video frame is greater than the preset first threshold include: Vfirst_1, Vfirst_2, Vfirst_3, Vfirst_4......Vfirst_n. Vfirst_1, Vfirst_2, Vfirst_3, Vfirst_4......Vfirst_n are added to the pseudo-label set and the reliable label set, respectively.

[0089] For each target reference frame in the reliable label set, determine whether any other video frame adjacent to the target reference frame is not a target reference frame. If the other video frame is not a target reference frame, compare the second confidence score corresponding to each object in the other video frame with a preset first threshold and a preset second threshold. If the second confidence score corresponding to any object in the other video frame is less than the preset first threshold and greater than the preset second threshold, and the other video frame is continuous with the target reference frame, then the other video frame is determined as a target reference frame and added to the pseudo-label set. The target reference frames contained in the pseudo-label set can be used to train the original video object detection model.

[0090] like Figure 3 As shown, the preceding video frame adjacent to Vfirst_1 is Vleft_1, and the following video frame adjacent to Vfirst_1 is Vright_1; the preceding video frame adjacent to Vfirst_2 is Vleft_2, and the following video frame adjacent to Vfirst_2 is Vright_2; the preceding video frame adjacent to Vfirst_3 is Vleft_3, and the following video frame adjacent to Vfirst_3 is Vright_3; the preceding video frame adjacent to Vfirst_4 is Vleft_4, and the following video frame adjacent to Vfirst_4 is Vright_4... the preceding video frame adjacent to Vfirst_n is Vleft_n, and the following video frame adjacent to Vfirst_n is Vright_n.

[0091] To increase the number of reliable target reference frames in the output results and facilitate subsequent improvements in the recognition accuracy of the video target detection model, based on the above embodiments, the method in this application embodiment further includes:

[0092] The other video frames are updated to the video frame, and the previous or next other video frame adjacent to the video frame is determined as the target reference frame until the other video frame is the target reference frame, or the second confidence of each object in the other video frame is less than the preset second threshold, or the other video frame is not continuous with the content of the video frame.

[0093] In this embodiment of the application, in order to increase the number of target reference frames with reliable output results, after determining that the output results of other video frames adjacent to the video frame are reliable and determining the other video as the target reference frame, it is also possible to determine whether the output results of video frames adjacent to the other video are reliable.

[0094] In this embodiment, after determining that the other video frame is the target reference frame, the other video is updated to a video frame, and the process continues to determine whether the preceding or following other video frame adjacent to the current video frame is the target reference frame, until the other video frame becomes the target reference frame, or the second confidence score of each object in the other video frame is less than a preset second threshold, or the content of the other video frame is not continuous with that of the current video frame. The process of determining whether the preceding and following other video frames adjacent to the updated video frame are target reference frames is the same as described above and will not be repeated here.

[0095] For example, a video may have 5 frames: frame 1, frame 2, frame 3, frame 4, and frame 5. If the second confidence level corresponding to any object in frame 3 is greater than a preset first threshold, then frame 3 is determined as the target reference frame. The second confidence level corresponding to each object in frames 1, 2, 4, and 5 is not greater than the preset first threshold, therefore frames 1, 2, 4, and 5 are non-target reference frames.

[0096] After determining video frame 3 as the target reference frame, based on the second confidence level of each object in the preceding and following video frames adjacent to video frame 3, it is determined whether the preceding and following video frames are target reference frames, that is, whether video frame 2 and video frame 4 are target reference frames. Since the second confidence level of each object in video frame 2 and video frame 4 is not greater than the preset first threshold, video frames 2 and 4 are not target reference frames. Therefore, it is determined whether the second confidence level of each object in video frame 2 and video frame 4 is less than the preset first threshold and greater than the preset second threshold, and whether the content is continuous with video frame 3. If it is determined that the second confidence level of any object in video frame 2 is less than the preset first threshold and greater than the preset second threshold, and the content is continuous with video frame 3, then video frame 2 is determined to be the target reference frame. The process continues to determine whether the preceding and following video frames adjacent to video frame 2 are target reference frames, i.e., whether video frame 1 and video frame 3 are target reference frames, until the other video frame is a target reference frame. That is, video frame 1 or video frame 3 is determined to be the target video frame, or the second confidence level of each object in the other video frame is less than the preset second threshold, or the content of the other video frame is not continuous with the target video frame.

[0097] If the second confidence level corresponding to each object in video frame 4 is less than the preset second threshold, or is not continuous with the content of video frame 3, then video frame 4 is a non-target reference frame. In this case, it is no longer necessary to determine whether the previous and next video frames adjacent to video frame 4 are target reference frames, that is, no operation is performed to determine whether video frame 3 and video frame 5 are target reference frames.

[0098] Since the operations for determining whether video frames 1 and 3 adjacent to video frame 2 are target reference frames, and the operations for determining whether video frames 2 and 4 adjacent to video frame 3 are target reference frames are the same as the above process, they will not be described in detail here.

[0099] Example 3:

[0100] To determine whether the content of two adjacent video frames is continuous, based on the above embodiments, in this embodiment, determining that the content of the other video frames is continuous with the content of the video frame includes:

[0101] Determine whether there is a target intersection-union ratio (IOU) greater than a preset IOU threshold among the intersection-union ratios (IOUs) between each object in the other video frames and each object in the video frame;

[0102] If so, then the other video frames are continuous with the content of the video frame.

[0103] To determine whether a video frame is content-continuous with other video frames, in this embodiment, an object detection algorithm can be used to determine the Intersection over Union (IOU) ratio between each object in the other video frame and each object in the current video frame. This IOU is then compared with a preset IOU threshold to determine whether the other video frame and the current video frame are content-continuous. Specifically, if there is a target IOU between each object in the other video frame and each object in the current video frame that is greater than the preset IOU threshold, then the other video frame and the current video frame are considered content-continuous.

[0104] Generally, the number of cases where the Intersection over Union (IOU) of two adjacent video frames is greater than 0.4 accounts for 99%. Therefore, in this embodiment, the preset IOU threshold can be 0.4. Specifically, the preset IOU threshold can be set according to requirements.

[0105] Example 4:

[0106] In order to update the parameters of the original video object detection model and obtain a trained video object detection model, based on the above embodiments, in this embodiment, updating the parameters of the original video object detection model according to the first loss value and the second loss value includes:

[0107] The parameters of the original video target detection model are updated based on the sum of the products of the first loss value, the second loss value, and their corresponding weight values.

[0108] In order to train the original video object detection model, in this embodiment of the application, a target loss value is determined based on a first loss value and a second loss value. Specifically, the target loss value can be determined based on the sum of the products of the first loss value, the second loss value and their corresponding weight values. The parameters of the original video object detection model are then updated based on the target loss value.

[0109] Figure 4 This is a schematic diagram illustrating the training process of a video object detection model provided in some embodiments of this application. Now, regarding... Figure 4 Line explanation.

[0110] The original video object detection model is trained based on the labeled image sample set. The first position information of each object in the sample image and the first confidence of the category to which each object belongs are obtained. Based on the first position information of each object in each sample image, the first confidence of the category to which each object belongs, the target position information of each object in the label of each sample image and the target confidence of the target category of each object, the first loss value is determined.

[0111] Input any video frame from the unlabeled video sample set into the original video object detection model to obtain the second location information of each object in each video frame output by the original video object detection model and the second confidence score of the category to which each object belongs.

[0112] For each video frame, determine the target reference frame.

[0113] The target reference frame is input into the original video object detection model, which outputs the third position information of each object in the target reference frame and the third confidence score of the category to which each object belongs. Based on the second position information and the second confidence score of the category to which each object belongs in the target reference frame, as well as the third position information and the third confidence score of the category to which each object belongs, a second loss value is determined.

[0114] The parameters of the original video object detection model are updated based on the sum of the products of the first loss value, the second loss value, and their corresponding weight values.

[0115] Figure 5 This is a schematic diagram of a target detection process provided in some embodiments of this application, and is now specifically for... Figure 2 Please provide an explanation.

[0116] First, train the original video object detection model based on the image sample set and the video sample set to obtain the trained video object detection model. After obtaining the trained video object detection model, directly deploy the trained video object detection model on electronic devices.

[0117] The electronic device obtains the video frame to be identified, inputs the video frame into a pre-trained video object detection model, obtains the position information of each object in the video frame to be identified and the confidence score of the category to which each object belongs, and outputs the position information of each object and the confidence score of the category to which each object belongs.

[0118] Specifically, the location information of each object and the confidence level of each object's category can be output based on voice or text. In other words, the location information of each object and the confidence level of each object's category can be played by voice, or the location information of each object and the confidence level of each object's category can be displayed on the display page by text.

[0119] Example 5:

[0120] Figure 6 This application provides a schematic diagram illustrating the effect of model training results in some embodiments. The diagram is now intended to... Figure 6 Please provide an explanation.

[0121] After obtaining the image sample set and the video sample set, the ratio of the number of sample images in the image sample set to the number of video samples in the video sample set can be any ratio. In this embodiment, a ratio of 1:9 is used for illustration. The specific training results are as follows: Figure 6 As shown.

[0122] When the ratio of the number of sample images in the image sample set to the number of video samples in the video sample set is 1:9, the baseline of the original video object detection model trained on the image sample set is 57.54, and the baseline of the original video object detection model trained on the video sample set is 63.37. Compared with the baseline of the original video object detection model trained only on the image sample set, the baseline increased by 5.83.

[0123] Training with a ratio of 1:9 between the number of sample images in the image sample set and the number of video samples in the video sample set surpasses the performance of training with twice the number of labeled sample images compared to training with only twice the number of sample images in the image sample set.

[0124] Example 6:

[0125] Figure 7 This application provides a schematic diagram of a video object detection model training device according to some embodiments. The device includes:

[0126] The acquisition module 701, for any sample image in a pre-saved image sample set, inputs the sample image into the original video object detection model to acquire the first position information of each object in the sample image and the first confidence score of the category to which each object belongs; for any video frame in a pre-saved video sample set, inputs the video frame into the original video object detection model to acquire the second position information of each object in the video frame and the second confidence score of the category to which each object belongs; if the second confidence score of any object is greater than a preset first threshold, the video frame is used as a target reference frame; the target reference frame is input into the original video object detection model to acquire the third position information of each object in the target reference frame and the third confidence score of the category to which each object belongs;

[0127] The determining module 702 is used to determine a first loss value based on the target location information of each object identified in the label of each sample image and the target confidence of the target category of each object, and the first location information and the first confidence of each object corresponding to each object; and to determine a second loss value based on the second location information and the second confidence of each object in the target reference frame, and the third location information and the third confidence of each object corresponding to each object.

[0128] The update module 703 is used to update the parameters of the original video target detection model based on the first loss value and the second loss value.

[0129] In one possible implementation, the acquisition module 701 is further configured to determine whether to identify the other video frames as target reference frames based on the second confidence level corresponding to each object in other video frames adjacent to the video frame and whether the other video frames are content-continuous with the video frame.

[0130] In one possible implementation, the acquisition module 701 is specifically used to determine the other video frame as the target reference frame if the other video frame adjacent to the video frame is not the target reference frame, the second confidence level corresponding to any object in the other video frame is less than the preset first threshold and greater than the preset second threshold, and the other video frame is continuous with the content of the video frame.

[0131] In one possible implementation, the acquisition module 701 is further configured to update the other video frames to the video frame, and continue to determine whether the previous or next other video frame adjacent to the video frame is the target reference frame, until the other video frame is the target reference frame, or the second confidence of each object in the other video frame is less than a preset second threshold, or the other video frame is not continuous with the content of the video frame.

[0132] In one possible implementation, the acquisition module 701 is specifically used to determine whether there is a target cross-union ratio (IOU) greater than a preset IOU threshold among the cross-union ratios (IOU) of each object in the other video frames and each object in the video frame; if so, the content of the other video frames and the video frame is continuous.

[0133] In one possible implementation, the update module 703 is specifically used to update the parameters of the original video target detection model based on the sum of the products of the first loss value, the second loss value, and their respective weight values.

[0134] Example 7:

[0135] Based on the above embodiments, some embodiments of this application also provide an electronic device, such as... Figure 8 As shown, it includes: processor 801, communication interface 802, memory 803 and communication bus 804, wherein processor 801, communication interface 802 and memory 803 communicate with each other through communication bus 804.

[0136] The memory 803 stores a computer program, which, when executed by the processor 801, causes the processor 801 to perform the following steps:

[0137] For any sample image in a pre-saved image sample set, the sample image is input into the original video object detection model to obtain the first position information of each object in the sample image and the first confidence level of the category to which each object belongs;

[0138] For any video frame in the pre-saved video sample set, the video frame is input into the original video target detection model to obtain the second position information of each object in the video frame and the second confidence score of the category to which each object belongs; if the second confidence score of any object is greater than a preset first threshold, the video frame is used as the target reference frame.

[0139] The target reference frame is input into the original video target detection model to obtain the third position information of each object in the target reference frame and the third confidence score of the category to which each object belongs;

[0140] A first loss value is determined based on the target location information of each object identified in the label of each sample image and the target confidence of the target category of each object, and the first location information and the first confidence of each object. A second loss value is determined based on the second location information and the second confidence of each object in the target reference frame, and the third location information and the third confidence of each object.

[0141] The parameters of the original video target detection model are updated based on the first loss value and the second loss value.

[0142] Furthermore, the processor 801 is also configured to determine whether to identify the other video frames as target reference frames based on the second confidence level corresponding to each object in other video frames adjacent to the video frame, and whether the other video frames are content-continuous with the video frame.

[0143] Furthermore, the processor 801 is also configured to, for any other video frame preceding or following the video frame, if the other video frame is not the target reference frame, the second confidence level corresponding to any object in the other video frame is less than the preset first threshold and greater than the preset second threshold, and the content of the other video frame is continuous with that of the video frame, then determine the other video frame as the target reference frame.

[0144] Furthermore, the processor 801 is also configured to update the other video frames to the video frame, and continue to determine whether the previous or next other video frame adjacent to the video frame is the target reference frame, until the other video frame is the target reference frame, or the second confidence of each object in the other video frame is less than a preset second threshold, or the other video frame is not continuous with the content of the video frame.

[0145] Furthermore, the processor 801 is also configured to determine whether there is a target cross-union ratio (IOU) greater than a preset IOU threshold among the cross-union ratios (IOU) between each object in the other video frames and each object in the target reference frame; if so, the content of the other video frames is continuous with that of the video frame.

[0146] Furthermore, the processor 801 is also configured to update the parameters of the original video target detection model based on the sum of the products of the first loss value, the second loss value, and the corresponding weight values.

[0147] The communication bus mentioned in the above server can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0148] Communication interface 802 is used for communication between the above-mentioned electronic device and other devices.

[0149] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0150] The processors mentioned above can be general-purpose processors, including central processing units, network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0151] Example 8:

[0152] Based on the above embodiments, this application also provides a computer-readable storage medium storing a computer program executable by an electronic device. When the program is run on the electronic device, the electronic device performs the following steps:

[0153] The memory stores a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0154] For any sample image in a pre-saved image sample set, the sample image is input into the original video object detection model to obtain the first position information of each object in the sample image and the first confidence level of the category to which each object belongs;

[0155] For any video frame in the pre-saved video sample set, the video frame is input into the original video target detection model to obtain the second position information of each object in the video frame and the second confidence score of the category to which each object belongs; if the second confidence score of any object is greater than a preset first threshold, the video frame is used as the target reference frame.

[0156] The target reference frame is input into the original video target detection model to obtain the third position information of each object in the target reference frame and the third confidence score of the category to which each object belongs;

[0157] A first loss value is determined based on the target location information of each object identified in the label of each sample image and the target confidence of the target category of each object, and the first location information and the first confidence of each object. A second loss value is determined based on the second location information and the second confidence of each object in the target reference frame, and the third location information and the third confidence of each object.

[0158] The parameters of the original video target detection model are updated based on the first loss value and the second loss value.

[0159] Furthermore, after using the video frame as the target reference frame, the method further includes:

[0160] Based on the second confidence level of each object in other video frames adjacent to the video frame, and whether the other video frames are content-continuous with the video frame, it is determined whether to identify the other video frames as target reference frames.

[0161] Further, determining whether to identify the other video frames as target reference frames based on the second confidence level corresponding to each object in other video frames adjacent to the video frame, and whether the other video frames are content-continuous with the video frame, includes:

[0162] For any other video frame preceding or following the video frame adjacent to the video frame, if the other video frame is not the target reference frame, the second confidence level corresponding to any object in the other video frame is less than the preset first threshold and greater than the preset second threshold, and the content of the other video frame is continuous with that of the video frame, then the other video frame is determined as the target reference frame.

[0163] Furthermore, the method also includes:

[0164] The other video frames are updated to the video frame, and the previous or next other video frame adjacent to the video frame is determined as the target reference frame until the other video frame is the target reference frame, or the second confidence of each object in the other video frame is less than the preset second threshold, or the other video frame is not continuous with the content of the video frame.

[0165] Further, determining that the other video frames are continuous with the content of the video frames includes:

[0166] Determine whether there is a target cross-union ratio (IOU) greater than a preset IOU threshold among the IOUs of each object in the other video frames and each object in the target reference frame;

[0167] If so, then the other video frames are continuous with the content of the video frame.

[0168] Further, updating the parameters of the original video target detection model based on the first loss value and the second loss value includes:

[0169] The parameters of the original video target detection model are updated based on the sum of the products of the first loss value, the second loss value, and their corresponding weight values.

[0170] In this embodiment, in addition to training the original video target detection model based on the input of labeled sample images from the image sample set, the original video target detection model is also trained based on unlabeled target reference frames from the video sample set with a confidence level greater than a preset first threshold. That is, this application adopts a semi-supervised learning method, which can train a high-precision video target detection model even without labeling all video frames, thereby improving the accuracy of target detection.

[0171] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0172] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0173] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0174] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0175] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for training a video object detection model, characterized in that, The method includes: For any sample image in a pre-saved image sample set, the sample image is input into the original video object detection model to obtain the first position information of each object in the sample image and the first confidence level of the object's category. For any video frame in the pre-saved video sample set, the video frame is input into the original video target detection model to obtain the second position information of each object in the video frame and the second confidence score of the category to which each object belongs; if the second confidence score of any object is greater than a preset first threshold, the video frame is used as the target reference frame. The target reference frame is input into the original video target detection model that is in the training process to obtain the third position information of each object in the target reference frame and the third confidence of the category to which each object belongs; A first loss value is determined based on the target location information of each object identified in the label of each sample image and the target confidence of the target category of each object, and the first location information and the first confidence of each object. A second loss value is determined based on the second location information and the second confidence of each object in the target reference frame, and the third location information and the third confidence of each object. The parameters of the original video target detection model are updated based on the first loss value and the second loss value.

2. The method according to claim 1, characterized in that, After using the video frame as the target reference frame, the method further includes: Based on the second confidence level of each object in other video frames adjacent to the video frame, and whether the other video frames are content-continuous with the video frame, it is determined whether to identify the other video frames as target reference frames.

3. The method according to claim 2, characterized in that, The step of determining whether to identify other video frames as target reference frames based on the second confidence level of each object in other video frames adjacent to the video frame, and whether the other video frames are content-continuous with the video frame, includes: For any other video frame preceding or following the video frame adjacent to the video frame, if the other video frame is not the target reference frame, the second confidence level corresponding to any object in the other video frame is less than the preset first threshold and greater than the preset second threshold, and the content of the other video frame is continuous with that of the video frame, then the other video frame is determined as the target reference frame.

4. The method according to claim 3, characterized in that, The method further includes: The other video frames are updated to the video frame, and the previous or next other video frame adjacent to the video frame is determined as the target reference frame until the other video frame is the target reference frame, or the second confidence of each object in the other video frame is less than the preset second threshold, or the other video frame is not continuous with the content of the video frame.

5. The method according to claim 2 or 3, characterized in that, Determining that the other video frames are continuous with the content of the video frame includes: Determine whether there is a target intersection-union ratio (IOU) greater than a preset IOU threshold among the intersection-union ratios (IOUs) between each object in the other video frames and each object in the video frame; If so, then the other video frames are continuous with the content of the video frame.

6. The method according to claim 1, characterized in that, The step of updating the parameters of the original video object detection model based on the first loss value and the second loss value includes: The parameters of the original video target detection model are updated based on the sum of the products of the first loss value, the second loss value, and their corresponding weight values.

7. A video object detection model training device, characterized in that, The device includes: The acquisition module is used to: input any sample image from a pre-saved image sample set into the original video object detection model to acquire the first position information and the first confidence score of the object's category for each object in the sample image; input any video frame from a pre-saved video sample set into the original video object detection model to acquire the second position information and the second confidence score of the object's category for each object in the video frame; if the second confidence score of any object is greater than a preset first threshold, the video frame is used as a target reference frame; input the target reference frame into the original video object detection model during training to acquire the third position information and the third confidence score of the object's category for each object in the target reference frame. The determination module is used to determine a first loss value based on the target location information of each object identified in the label of each sample image and the target confidence of the target category of each object, and the first location information and the first confidence of each object; and to determine a second loss value based on the second location information and the second confidence of each object in the target reference frame, and the third location information and the third confidence of each object. An update module is used to update the parameters of the original video target detection model based on the first loss value and the second loss value.

8. The apparatus according to claim 7, characterized in that, The acquisition module is further configured to determine whether to identify the other video frames as target reference frames based on the second confidence level of each object in other video frames adjacent to the video frame and whether the other video frames are content-continuous with the video frame.

9. The apparatus according to claim 8, characterized in that, The acquisition module is specifically used to determine the other video frame as the target reference frame if the other video frame is not the target reference frame, the second confidence level of any object in the other video frame is less than the preset first threshold and greater than the preset second threshold, and the other video frame is continuous with the content of the video frame.

10. The apparatus according to claim 9, characterized in that, The acquisition module is further configured to update the other video frames to the video frame, and continue to determine whether the previous or next other video frame adjacent to the video frame is the target reference frame, until the other video frame is the target reference frame, or the second confidence of each object in the other video frame is less than a preset second threshold, or the other video frame is not continuous with the content of the video frame.

11. The apparatus according to claim 8 or 9, characterized in that, The acquisition module is specifically used to determine whether there is a target cross-union ratio (IOU) greater than a preset IOU threshold among the intersection-union ratios (IOUs) between each object in the other video frames and each object in the target reference frame; if so, the content of the other video frames is continuous with that of the video frame.

12. The apparatus according to claim 7, characterized in that, The update module is specifically used to update the parameters of the original video target detection model based on the sum of the products of the first loss value, the second loss value, and their respective weight values.

13. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory being used to store program instructions, and the processor being used to execute the computer program stored in the memory to implement the steps of the video target detection model training method according to any one of claims 1-6.

14. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the steps of the video target detection model training method according to any one of claims 1-6.

Images