Labeling of perception information, evaluation method and device of detection model, and storage medium

Abstract

The present application relates to the technical field of automatic driving, and particularly provides a kind of perception information's marking, detection model's evaluation method, device and storage medium, to solve how to more efficient, low-cost acquisition problem perception information's marking result, and the test method of the problem of constructing perception information detection model.For this purpose, the present application constructs evaluation dataset according to the problem perception information to be marked, according to the corresponding perception problem type of evaluation demand, the problem perception information in evaluation dataset is regionally marked and / or target marked, obtains marking result, can avoid the process of true value marking to all targets of evaluation dataset, realizes more quickly and simply the marking process of problem perception information, both can directly obtain the required marking result, and can reduce marking cost.At the same time, the evaluation of perception information detection model is carried out by using the obtained marking result, and the evaluation efficiency of the perception information detection model can also be effectively improved.

Claims

1. A method for labeling perceived information, characterized in that, The method includes: Obtain problem-aware information to be labeled, and construct an evaluation dataset using the problem-aware information; Based on the perceived problem type corresponding to the evaluation requirements, the perceived problem information in the evaluation dataset is selectively labeled using region labeling and / or target labeling to obtain the labeling results of the perceived problem information; The evaluation dataset includes visualized data; "Constructing the evaluation dataset using the problem-aware information" includes: The problem-aware information is transformed to obtain visualized data of the problem-aware information. The evaluation dataset is constructed based on the visualization data. The steps for "adding annotations" include: Based on the annotation result of one of the adjacent frames of the visualization data, the annotation result of the other frame in the adjacent frame is obtained.

2. The annotation method for perceived information according to claim 1, characterized in that, The perception problem types include at least the false alarm problem and / or missed alarm problem of the preset target; "Selectively applying region labeling and / or target labeling to the problem-aware information in the evaluation dataset" includes: Based at least on the false alarm and / or missed alarm issues of the preset target, the regions in the problem perception information that do not contain the preset target are labeled.

3. The method for labeling perceived information according to claim 1 or 2, characterized in that, The types of perception problems include at least the splitting problem and / or inaccuracy problem of the preset target; "Selectively applying region labeling and / or target labeling to the problem-aware information in the evaluation dataset" also includes: At least based on the splitting problem and / or inaccuracy problem of the preset target, the preset target in the problem perception information is labeled.

4. A method for evaluating a perceptual information detection model, characterized in that, The method includes: Obtain a labeled dataset of problem-aware information; the problem-aware information in the labeled dataset is labeled based on the labeling method of the perception information according to any one of claims 1 to 3; The problem perception information is detected based on the perception information detection model to obtain the model perception result; Based on the model's perception results and the annotation results of the problem perception information in the labeled dataset, the evaluation results of the perception information detection model are obtained.

5. The evaluation method for the perceptual information detection model according to claim 4, characterized in that, The method further includes: Based on the evaluation results, determine the problems existing in the perceptual information detection model and the causes of the problems, and adjust the parameters of the perceptual information detection model according to the causes; and / or, The proportion of training data for the perceptual information detection model is adjusted based on the evaluation results; and / or, When there are multiple perception information detection models, the perception information detection model that meets the preset requirements is selected from among the multiple perception information detection models according to the evaluation results.

6. The evaluation method for the perceptual information detection model according to claim 4, characterized in that, The method further includes: Obtain the evaluation results of the current version and historical versions of the perception information detection model, respectively; The evaluation results of the current version are compared with those of historical versions to obtain the evaluation comparison results of the perceptual information detection model.

7. A control device comprising at least one processor and at least one storage device, said storage device being adapted to store a plurality of program codes, characterized in that, The program code is adapted to be loaded and run by the processor to perform the annotation method for the perceptual information according to any one of claims 1 to 3 or the evaluation method for the perceptual information detection model according to any one of claims 4 to 6.

8. A computer-readable storage medium storing a plurality of program codes, characterized in that, The program code is adapted to be loaded and run by a processor to perform the annotation method for the perceptual information as described in any one of claims 1 to 3 or the evaluation method for the perceptual information detection model as described in any one of claims 4 to 6.

Images