Data annotation quality inspection method

Abstract

The invention provides a data annotation quality inspection method. The method comprises the following steps: step 1, importing image data, point cloud data and joint data; step 2, detecting image frames of the image data, the point cloud data and the joint data and counting according to category names of the image data and the joint data as well as label names and data exchange files of the point cloud data; step 3, carrying out super-resolution reconstruction on all data; and step 4, importing the number of the image frames and the reconstructed data into the original data for quality inspection personnel to judge and feed back. The technical problems that in the prior art, data quality inspection completely depends on manual work, the accuracy of the quality inspection result cannot be guaranteed, and the quality inspection process consumes long time are solved.

Description

Technical Field

[0001] This application relates to the technical field of data annotation, and particularly to a data annotation quality inspection method. Background Art

[0002] The quality inspection work in the annotation task refers to checking and evaluating the annotation results to ensure the accuracy, consistency, and standardization of the annotated data. The quality inspection work mainly checks the annotated data to determine whether the annotation results are consistent with the actual situation and whether there are errors or deviations. The consistency check mainly examines whether the annotation results of different annotators for the same object are consistent and whether there are situations such as missing annotation or misannotation of the same object in the front and back frames. Currently, the data quality inspection work mainly relies on manual sampling inspection of the annotated data. By looking at the position and attributes of the annotated objects, it is judged whether the sampling results are correctly annotated. Missing annotation, over-annotation, and misannotation all belong to annotation errors. However, the efficiency of manual quality inspection for judging annotation situations is low, and the quality inspection results depend on the sampled data and the accuracy of the quality inspector's judgment. Therefore, a large amount of manpower is required for the quality inspection effect, and it is impossible to ensure the accuracy of the quality inspection and it takes a lot of time. Summary of the Invention

[0003] In view of this, the present invention provides a data annotation quality inspection method to solve the technical problems in the prior art that completely rely on manual data quality inspection, cannot guarantee the accuracy of the quality inspection results, and the quality inspection process takes a long time.

[0004] The present invention provides a data annotation quality inspection method, and the method includes: Step 1, import image data, point cloud data, and combined data; Step 2, detect the image frames of the image data, point cloud data, and combined data and count according to the category names of the image data and combined data, and the label names and data exchange files of the point cloud data; Step 3, perform super-resolution reconstruction on all data; Step 4, import the number of image frames and the reconstructed data into the original data for quality inspectors to judge and give feedback.

[0005] Further, the image data includes object detection data, instance segmentation data, traffic light data, and drivable area data.

[0006] Further, the method for detecting the image frames of the image data and counting includes: Step 211, query the annotations in the object detection data, instance segmentation data, and traffic light data; Step 212, obtain the category coding names in the annotations; Step 213, count according to the coding names; Step 214, obtain the category coding names in the drivable area data and count according to the coding names.

[0007] Further, the traffic light data includes a timing board, a lamp cluster, and a lamp frame. Among them, the timing board includes a digital panel, the lamp cluster includes suspended traffic lights, mobile traffic lights, and pole-mounted traffic lights, and the lamp frame includes lamp heads.

[0008] Further, the point cloud data includes target detection data and semantic segmentation calculation data.

[0009] Further, the method for detecting and counting the image frames of the detection point cloud data includes: Step 221, querying the annotations in the target detection data; Step 222, obtaining the category encoding names in the annotations and counting according to the encoding names; Step 223, counting and summarizing the data exchange files included in each corresponding path in the semantic segmentation calculation data.

[0010] Further, the combined data includes target detection data, top-down target detection data, and other target detection data.

[0011] Further, the method for detecting and counting the image frames of the combined data includes: Step 231, obtaining the category encoding names in the annotations of the top-down target detection data and counting according to the encoding names; Step 232, obtaining the label names in the annotations of the target detection data and counting according to the label names; Step 233, obtaining the label names in the annotations of the other target detection data and counting according to the label names.

[0012] Further, the method further includes: Step 5, labeling the attributes of all image data and separating them by left and right sides; Step 6, respectively detecting the image data on both sides and automatically labeling data anomalies according to the labeled attributes.

[0013] Further, Step 6 includes: Step 61, respectively detecting the image data on both sides; Step 62, sequentially labeling the attribute serial numbers of the unilateral image data from far to near; Step 63, when there is a jump or equality of the serial numbers, automatically labeling the data as abnormal.

[0014] The present invention provides a method for data annotation quality inspection. This method detects image frames and then performs super-resolution reconstruction, which is more conducive to quality inspectors observing details. Then, the processed data is compared with manual judgment, and it is determined whether the data annotation is incorrect by counting the categories and quantities of the annotation frames. This method no longer completely relies on manual judgment, improves the quality inspection accuracy, and also saves quality inspection time. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic flow chart of a method for data annotation quality inspection provided by the present invention; Figure 2 is a schematic flow chart of a method for detecting and counting the image frames of the detection image data provided by the present invention; Figure 3 It is a schematic flowchart of the method for detecting and counting the image frames of point cloud data provided by the present invention; Figure 4 It is a schematic flowchart of the method for detecting and counting the image frames of combined data provided by the present invention; Figure 5 It is a schematic flowchart of another data annotation quality inspection method provided by the present invention; Figure 6 It is a schematic diagram of the interface for separating the left and right sides of the image annotation attributes provided by the present invention. Detailed implementation manners

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0017] Embodiment 1: The present invention provides a data annotation quality inspection method, which provides a data annotation quality inspection method for the situation that a large amount of human resources of quality inspectors are required for sampling inspection in the prior art. As Figure 1 shown, the method includes the following steps.

[0018] Step 1, import image data, point cloud data, and combined data; Step 2, detect the image frames of the image data, point cloud data, and combined data and count them according to the category names of the image data and combined data, as well as the label names and data exchange files of the point cloud data; Step 3, perform super-resolution reconstruction on all the data; Step 4, import the number of image frames and the reconstructed data into the original data for quality inspectors to make judgments and give feedback.

[0019] The present invention provides a data annotation quality inspection method. This method is more conducive to quality inspectors observing details by detecting image frames and then performing super-resolution reconstruction, and then comparing the processed data with manual judgment to determine whether the data annotation is incorrect by counting the category and quantity of the annotation frames. This method no longer completely relies on manual judgment, improves the quality inspection accuracy, and also saves the quality inspection time.

[0020] Embodiment 2: The present invention provides a data annotation quality inspection method. As Figure 1 shown, the method includes the following steps.

[0021] Step 1, import image data, point cloud data, and combined data; Step 2: Detect the image frames of the image data, point cloud data, and combined data according to the class names of the image data and combined data, as well as the label names and data exchange files of the point cloud data, and count them. The following will introduce in detail the process of detecting and counting the image frames of these three types of data: image data, point cloud data, and combined data.

[0022] The first is the method steps for detecting and counting the image frames of the image data. The image data includes object detection data, instance segmentation data, traffic light data, and drivable area data. As Figure 2 shown, the method for detecting and counting the image frames of the image data includes the following steps.

[0023] Step 211: Query the annotations in the object detection data, instance segmentation data, and traffic light data. The traffic light data includes a timing board, a lamp cluster, and a lamp frame. Among them, the timing board includes a digital panel, the lamp cluster includes a suspended traffic light, a movable traffic light, and a pole-mounted traffic light, and the lamp frame includes a lamp head.

[0024] Step 212: Obtain the class encoding names in the annotations. Step 213: Count according to the encoding names. Step 214: Obtain the class encoding names in the drivable area data and count according to the encoding names.

[0025] The second is the method steps for detecting and counting the image frames of the point cloud data. The point cloud data includes object detection data and semantic segmentation calculation data. As Figure 3 shown, the method for detecting and counting the image frames of the point cloud data includes the following steps.

[0026] Step 221: Query the annotations in the object detection data. Step 222: Obtain the class encoding names in the annotations and count according to the encoding names. Step 223: Count and summarize the data exchange files included in each corresponding path in the semantic segmentation calculation data.

[0027] The third is the method steps for detecting and counting the image frames of the combined data. The combined data includes object detection data, top-down object detection data, and other object detection data. As Figure 4 shown, the method for detecting and counting the image frames of the combined data includes the following steps.

[0028] Step 231: Obtain the class encoding names in the annotations of the top-down object detection data and count according to the encoding names. Step 232: Obtain the label names in the target detection data annotation and count according to the label names; Step 233: Obtain the label names in other target detection data annotations and count according to the label names.

[0029] In the above content, the category coding names (the names of category IDs), label names, and data exchange files (json files, a data exchange format using JavaScript Object Notation, mainly used to store and exchange text information. Its syntax is simple, easy to read and write, and also easy for machines to parse and generate. The extension of JSON files is usually.json, and it can exchange data between different programming languages, so it is very common in web development) can all be used as image frames and counted. Image frames are used to display graphic information. Controls use fewer system resources and have a fast redrawing speed, and can extend the size of the picture to fit the size of the control.

[0030] Step 3: Perform super-resolution reconstruction on all data; Step 4: Import the number of image frames and the reconstructed data into the original data for quality inspection personnel to make judgments and give feedback.

[0031] The present invention provides a data annotation quality inspection method. This method detects image frames and counts, then performs super-resolution reconstruction, which is more conducive for quality inspection personnel to observe details. Then, the processed data is compared with manual judgment, and it is determined whether the data annotation is incorrect by counting the categories and quantities of annotation frames. This method no longer completely relies on manual judgment, improves the quality inspection accuracy, and also saves quality inspection time.

[0032] Embodiment 3: The present invention provides a data annotation quality inspection method, as Figure 5 shown. The method includes the following steps.

[0033] Step 1: Import image data, point cloud data, and combined data; Step 2: Detect the image frames of the image data, point cloud data, and combined data and count according to the category names of the image data and combined data, as well as the label names and data exchange files of the point cloud data; Step 3: Perform super-resolution reconstruction on all data; Performing super-resolution reconstruction on the original data can improve the image resolution of the original picture, improve the picture quality observed by the quality inspector, and facilitate the quality inspector to better identify the target annotation situation.

[0034] Step 4: Import the number of image frames and the reconstructed data into the original data for quality inspectors to make judgments and provide feedback.

[0035] Import all the read data into the original images, point clouds, and combined data for visual display. The quality inspector only needs to quickly check whether the positions and attributes of the annotation data provided by the data annotator are consistent with manual judgment. If they are inconsistent, record them and feedback to the data annotator for re-annotation, greatly reducing the time required for data quality inspection.

[0036] Step 5: Separate all the image data annotation attributes by left and right sides; Step 6: Detect the image data on both sides respectively, and automatically mark the data as abnormal according to the annotated attributes.

[0037] The said Step 6 includes: Step 61: Detect the image data on both sides respectively; Step 62: Sequentially mark the attribute numbers of the unilateral image data from far to near; Step 63: When there is a jump or equality in the numbers, automatically mark the data as abnormal.

[0038] At the same time, the regularization process can separate the left and right sides of the image annotation attributes. For example, for the left side, the lane and drivable area annotation attributes are sequentially 1, 2, 3, etc. from near to far. If there is a jump or equality in the numbers, automatically mark the data as abnormal. As Figure 6 shown, the left lane is marked as 1, the right lane is marked as -1, and the opposite is 0. There is no jump or equality in the numbers, indicating that the annotation is normal.

[0039] The present invention provides a data annotation quality inspection method. This method detects the image frames and technology, then performs super-resolution reconstruction, which is more conducive to quality inspectors observing details. Then, the processed data is compared with manual judgment, and it is determined whether the data annotation is incorrect by counting the categories and quantities of the annotation frames. This method no longer completely relies on manual judgment, improves the quality inspection accuracy, and also saves the quality inspection time.

[0040] In summary, the embodiments of the present invention provide a data annotation quality inspection method. This method provides visual display of the data, and uses super-resolution reconstruction technology to improve the resolution of the original picture image, facilitating the quality inspector to observe the details of the picture contour. Then, combined with the manually annotated data, the annotation situation of the data can be quickly screened. Verify the manually annotated data, calculate the categories and quantities of the annotation frames, which is convenient for checking the standard qualification rate of the calculated data, and solves the problems that manual quality inspection requires a large amount of manpower, cannot ensure the accuracy of quality inspection, and will consume a lot of time.

[0041] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A data annotation quality inspection method, characterized in that: The method comprises: Step 1, import image data, point cloud data and joint data; Step 2: Detect and count the image frames of the image data, point cloud data, and joint data according to the category names of the image data and joint data, as well as the label names and data exchange files of the point cloud data; Step 3, super-resolution reconstruction of all data; Step 4: Import the number of image frames and the reconstructed data into the original data for quality inspectors to make judgments and provide feedback.

2. A data annotation quality inspection method according to claim 1, characterized in that: The image data includes target detection data, instance separation data, traffic light data, and drivable area data.

3. A data annotation quality inspection method according to claim 2, characterized in that: The method for detecting and counting image frames of image data comprises: Step 211, querying annotations in the target detection data, instance separation data, and traffic light data; Step 212, obtaining the category code name in the annotation; Step 213, counting according to the code name; Step 214, obtaining the category code name in the drivable area data, and counting according to the code name.

4. A data annotation quality inspection method according to claim 3, characterized in that: The traffic light data includes a timing board, a light cluster, and a light frame, wherein the timing board includes a digital panel, the light cluster includes a suspended traffic light, a movable traffic light, and a staked traffic light, and the light frame includes a lamp head.

5. According to claim 1, a data annotation quality inspection method is characterized in that: The point cloud data includes target detection data and semantic segmentation calculation data.

6. A data annotation quality inspection method according to claim 5, characterized in that: The method for detecting and counting image frames of point cloud data comprises: Step 221, querying annotations in target detection data; Step 222, obtaining the category code name in the annotation, and counting according to the code name; Step 223: Count and summarize the data exchange files contained in each corresponding path in the semantic separation calculation data.

7. A data annotation quality inspection method according to claim 1, characterized in that: The combined data includes target detection data, overhead target detection data, and other target detection data.

8. A data annotation quality inspection method according to claim 7, characterized in that: The method for detecting and counting image frames of joint data comprises: Step 231, obtaining the category code name in the overhead target detection data annotation, and counting according to the code name; Step 232, obtaining the label name in the target detection data annotation, and counting according to the label name; Step 233, obtain the label names in the annotations of other target detection data, and count according to the label names.

9. A data annotation quality inspection method according to claim 1, characterized in that: The method further comprises: Step 5: annotate all image data with attributes and separate them into left and right sides; Step 6: Detect the image data on both sides respectively, and automatically mark the data anomalies according to the marked attributes.

10. A data annotation quality inspection method according to claim 9, characterized in that: The step 6 comprises: Step 61, detecting the image data on both sides respectively; Step 62, the single-side image data is labeled with attribute numbers in order from far to near; Step 63, when the sequence number jumps or is the same, the data is automatically marked as abnormal.

Images