Automatic model evaluation method and device based on batch normalization layer

Abstract

The invention discloses an automatic model evaluation method and device based on a batch normalization layer. The method comprises the following steps: after first training is completed, extracting batch normalization layer parameters of a fourth module of a model as a measurement reference; training the model for the second time on a label-free test data set needing to be evaluated, wherein in the second training process, parameters of a non-batch normalization layer do not need to be updated, gradient back propagation is not needed, and only the two parameters of the mean value and the variance of the batch normalization layer need to be updated; after the second training is completed, extracting batch normalization layer parameters of the fourth module of the model after the second training; respectively carrying out Gaussian normalization on the reference parameters obtained by the first training and the parameters extracted after the second training on the new test set, then calculating the Euclidean distance between the two groups of parameters, and using the distance to evaluate the precision of the label-free test set on the original model.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to a model automatic evaluation method and device based on a batch normalization layer. Background technique [0002] Target re-identification includes pedestrian re-identification and vehicle re-identification. It aims to retrieve the user-specified pedestrian or vehicle from a series of cross-camera surveillance videos. This technology is widely used in smart cities, security monitoring and other fields. [0003] To solve object re-identification tasks, deep learning-based methods usually require datasets to train and test the models. The one used to train the model is called the training set, and the one used to test the accuracy of the model is called the test set. In order to calculate the accuracy of a model, a test set consisting of test sample images and their true labels is usually required. Although standard test cases can be used to meet this requirement, in a...

AI Technical Summary

Problems solved by technology

Although this requirement can be met by using standard test cases, in actual application scenarios, the sample size of test data is very large, and it will consume a lot of manpower and material resources if the data is manually labeled

Therefore, many real-world scenarios contain unlabeled test data, making general model evaluation methods infeasible, thus triggering the research of automatic model evaluation methods.

Images