Unsupervised Domain Adaptive Object Detection Based on Frequency Domain Adjustment and Pixel-Level Feature Fusion
Yanlong Xu, Huijie Fan, Hao Pan, Lianquan Wu, Yandong Tang · 2022 12th International Conference on CYBER Technology in Automation, Control, and Intelligent Systems (CYBER) · 2022
It is very risky to apply the knowledge learned by the model on the labeled dataset directly to a completely new dataset, especially the unlabeled data, as the gap between these two domains can lead to a severe drop in model performance, However, the difference in illumination, texture and background of the image is one of the difficult problems to solve. In this paper, We proposed a Frequency domain Adjustment module (FA) and Pixel-Level feature Fusion (PLF) block for unsupervised domain adaptation of different types of features. From a frequency domain perspective, we first transfer the style of unlabeled target domain data to labeled source domain data before training using a Fourier transform. The benefit of this strategy can narrow the illumination gap between the source domain and the target domain. Secondly, due to the lack of labels in the target domain data, the feature extraction effect is poor. We consider proposing an improved multi-scale feature fusion mechanism to integrate information. Our method is experimentally validated on multiple public adaptive public datasets, The experimental results show that this method can more accurately detect and recognize images in domain adaptive scenes, and it outperforms the mainstream methods of domain adaptive object detection.