Collaborative training method based on domain self-adaptation

Abstract

The invention discloses a collaborative training method based on domain self-adaptation, which comprises the following steps: training one party by utilizing high confidence output of the other party,and processing candidate regions with low confidence output of the two parties by utilizing a maximized classifier difference method; besides, in the aspect of feature alignment of the backbone network, the output of the RPN is used for calculating the foreground probability of each point on the feature map, and a larger weight is given to an area with a larger foreground probability during feature alignment; the target detection capability of the model in the label-free field is improved, the requirement of the target detection model for annotation data is reduced, and dependence on human resources is reduced.

Description

technical field [0001] The invention relates to the field of model training, in particular to a collaborative training method based on domain self-adaptation. Background technique [0002] The development of deep learning has made great progress in the field of computer vision. However, the huge amount of labeled data required for deep learning model training restricts the wide application of deep learning. Object detection tasks require more fine-grained annotations, so the problem is particularly serious. Domain adaptation tries to solve this problem by transferring the model from a domain rich in labeled data (source domain) to a domain lacking in labeled data (target domain). [0003] At present, domain adaptation in object detection is mainly divided into two types of schemes. [0004] (1) Domain adaptation at the feature level, the main method is to achieve the alignment of two domain features through adversarial learning. At this time, since the image of the target...

AI Technical Summary

Problems solved by technology

[0009] When transferring Faster RCNN to an unlabeled target domain,

Images