SR-DAYOLOv8: cross-domain adaptive object detection based on super-resolution domain classifier: SR-DAYOLOv8: cross-domain adaptive object detection

Object detection is a fundamental task of environment perception in traffic road scenarios, and its accurate detection results are of great significance for improving the reliability of autonomous driving, optimizing traffic flow management, and enhancing road safety. However, the problem of domain...

Full description

Saved in:
Bibliographic Details
Published in:Multimedia systems 2025, Vol.31 (1)
Main Authors: Wang, Huilin, Qian, Huaming
Format: Article
Language:English
Subjects:
Computer Communication Networks
Computer Graphics
Computer Science
Cryptology
Data Storage Representation
Multimedia Information Systems
Operating Systems
Regular Paper
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Object detection is a fundamental task of environment perception in traffic road scenarios, and its accurate detection results are of great significance for improving the reliability of autonomous driving, optimizing traffic flow management, and enhancing road safety. However, the problem of domain offset between different traffic road scenarios leads to a poor generalization of the target detector. To address this challenge, we propose a new unsupervised domain adaptation object detection algorithm, SR-DAYOLOv8. Specifically, the algorithm contains three effective components. First, we use the source and target domains to train an unpaired image-to-image translator to generate a target-like domain, using the target-like domain as input to compensate for image-level differences. Second, to correct cross-domain discrepancies, we add a new detector for the target-like domain, enabling it to conduct supervised learning training, just like the source domain. Finally, we design a super-resolution domain classifier to obtain domain adaptive feature maps. Domain-invariant features are extracted through image-level adaptation and instance-level adaptation, and consistency regularization is employed to optimize the overall alignment effect. We conducted experiments on Cityscape, Foggy Cityscape, KITTI, SIM10K, and BDD100K datasets for different domain offset scenarios. Experimental results show that our method can improve target detection performance in different domain offset scenarios and outperform other state-of-the-art algorithms.
ISSN:0942-4962
1432-1882
DOI:10.1007/s00530-024-01594-4