Prediction result identification method, prediction result identification model training method, prediction result identification model training device and computer storage medium

Abstract

A prediction result identification method, a prediction result identification model training method, a prediction result identification model training device and a computer storage medium mainly comprise the steps of determining a category label and a position label of an annotation box in a sample picture according to the sample picture, obtaining each anchor frame prediction result of each anchor frame in the sample picture, and obtaining each anchor frame prediction result of each anchor frame; according to each category prediction information and each position prediction information of each anchor frame, and according to the category label and the position label, determining at least one anchor frame prediction result to be identified as a first sample; according to the position prediction information of each anchor frame and the category label and the position label of the labeling frame, determining at least one anchor frame prediction result to be identified as a second sample. Therefore, better positive and negative sample distribution reference can be obtained, and the training performance of the model can be optimized.

Description

technical field [0001] The embodiments of the present application relate to the technical field of sample identification, and more specifically, to a prediction result identification and its model training method, device and computer storage medium. Background technique [0002] In the current target detection based on deep learning, the distribution of positive and negative samples is generally based on the intersection ratio of the predefined anchor point and the real target frame. A positive threshold and a negative threshold are set. When the intersection of the anchor point and the real target frame When the union ratio is greater than the positive threshold, the sample is a positive sample, and when it is smaller than the negative threshold, the sample is a negative sample. [0003] However, there are two shortcomings in this positive and negative sample distribution method: first, it depends on the design of the anchor point, and the design of the anchor point is pron...

AI Technical Summary

Problems solved by technology

[0003] However, there are two shortcomings in this positive and negative sample distribution method: first, it depends on the design of the anchor point, and the design of the anchor point is prone to blunt shortcomings due to manual generation, for example, the positive sample that is easy to cause large targets to be matched Many and small targets are matched with fewer positive samples; second, when assigning positive samples, due to the existence of occlusion, the center of the anchor point assigned to a real target frame may actually fall on another target

However, this distribution method of positive and negative samples still has the following disadvantages: first, a large number of low-quality positive and negative samples are introduced in the training process, which is not conducive to the optimization of the network; second, the loss weight of negative samples is reduced, which may increase false positive risk

Images