Affiliation:
1. Zhongshan Institute University of Electronic Science and Technology Zhongshan China
2. College of Mathematics and Informatics College of Software Engineering South China Agricultural University Guangzhou China
Abstract
AbstractCross‐domain object detection aims to generalize the distribution of features extracted by an object detector from an annotated domain to an unknown and unlabelled domain. Although one‐stage cross‐domain object detectors have significant advantages in deployment than two‐stage ones, they suffer from two problems. First, neglect of category features and inaccurate alignment between multiple category features would lead to decreased domain adaptation efficiency. Second, one‐stage detectors are more sensitive to imbalance of samples and negative samples severely affect the alignment process of domain adaptation. To overcome these two problems, an innovative category‐related attention domain adaptive method that refines discrimination for each category's feature has been proposed in this paper. In the proposed method, a group of domain discriminators is assigned to each category to refine the fine‐grained features between categories. The discriminators are trained via an adversarial discriminant framework to align the fine‐grained distributions cross different domains. A category attention alignment (CAA) module is proposed to navigate more attention to the foreground regions in instance‐level, which effectively alleviates the negative migration problem caused by the positive and negative sample imbalance of the one‐stage detector. Specifically, two sub‐modules in the CAA module are developed: a local CAA module and a global CAA module. These modules aim to optimize the domain offsets in both the local and global dimensions. In addition, a progressive global alignment module is designed to align image‐level features, offering prior knowledge of migration for the CAA module. The progressive global alignment module and CAA module collaboratively engage in benign competition with the backbone network across various levels. Extensive transferring experiments are conducted among cityscapes, foggy cityscapes, SIM10K, and KITTI. Experimental results show that the proposed method has much superior performance than other one‐stage cross‐domain detectors.
Funder
Natural Science Foundation of Guangdong Province
Medical Science and Technology Foundation of Guangdong Province
National Natural Science Foundation of China
Publisher
Institution of Engineering and Technology (IET)
Subject
Electrical and Electronic Engineering,Computer Vision and Pattern Recognition,Signal Processing,Software
Reference60 articles.
1. Girshick R.:Fast R‐CNN. In:2015 IEEE International Conference on Computer Vision (ICCV) pp.1440–1448.IEEE Piscataway NJ(2015)
2. Ren S. He K. Girshick R. Sun J.:Faster R‐CNN: towards real‐time object detection with region proposal networks. In:Advances in Neural Information Processing Systems pp.91–99.ACM New York(2015)
3. Redmon J. Divvala S. Girshick R. Farhadi A.:You only look once: unified real‐time object detection. In:2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp.779–788.IEEE Piscataway NJ(2016)
4. Redmon J. Farhadi A.:YOLOv3: an incremental improvement. arXiv:180402767 (2018)
5. Bochkovskiy A. Wang C.Y. Liao H.Y.M.:YOLOv4: optimal speed and accuracy of object detection. arXiv:200410934 (2020)