Cross-domain pedestrian re-identification algorithm based on momentum network guidance

Abstract

The invention provides a cross-domain pedestrian re-identification algorithm based on momentum network guidance in order to solve the problem of false label noise interference caused by a domain offset phenomenon existing in a cross-domain pedestrian re-identification task. The method comprises the following steps: s1, initializing a backbone network by using a pre-trained model on an ImageNet data set; s2, carrying out pre-fine tuning on the model by utilizing marked data on the source domain data set so as to fully utilize marked information of the source domain; s3, initializing a proposed momentum learning framework by using a model trained by setting different random parameters on the source domain data set, and clustering according to features extracted by the model by using a clustering algorithm to generate a hard pseudo tag with the confidence coefficient of 1; s4, designing a new softened pseudo tag and a loss function to be combined with the traditional loss to train an optimization model; s5, updating the hard pseudo tag before each round of iteration is started, dynamically updating the soft pseudo tag in real time, and continuously iterating to generate and optimize the pseudo tag until the model converges.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a cross-domain pedestrian re-identification algorithm guided by a momentum network. Background technique [0002] The task of pedestrian re-identification is given a target image, and it is necessary to find one or several images that are closest to the target image from the pedestrian database by some method. In recent years, the disclosure of many large-scale pedestrian datasets has promoted the research of pedestrian re-identification technology. It can be said that the quality and scale of the dataset are the key to improving the performance of pedestrian re-identification technology. However, the labeling of datasets is very labor-intensive, and environmental factors also hinder the collection of effective data to a certain extent. Considering that in practical applications, the robustness of the model directly determines its practicability. However, the study found that if a...

AI Technical Summary

Problems solved by technology

[0005] Purpose of the invention: The present invention aims at the problem of pseudo-label noise interference caused by domain offset, unknown number of pedestrian categories in the target domain, and limitations of the algorithm itself in the unsupervised pedestrian re-identification task, and proposes a cross-domain algorithm based on momentum network guidance. Pedestrian re-identification method to reduce pseudo-label noise in cross-domain scenarios and improve retrieval performance of unsupervised domain-adapted pedestrian re-identification model

Images