Unsupervised domain adaptation (UDA) suffers from misleading by misclassified samples when aligning the labeled source domain and the unlabeled target domain, which causes unreliable alignment and class overlap. In this paper, we propose a center-aware UDA algorithm for inter-domain alignment, along with class diversity regularization for target intra-domain separability. First, we align the class centers of the source and target domain by weighting the sample confidence to corresponding centers, to denoise the mislabeled samples. Second, we scatter the predictions of the target domain by minimizing the entropy of intra-class predictions and decreasing the similarity of inter-class distributions, which tackles the class overlap. The proposed method not only aligns the two domains but also increases the separability of the target domain, which results in improved target data classification performance. We validate the proposed method on behavior recognition and object recognition datasets, and the experimental results show that the method greatly alleviates the domain shift compared to the latest UDA methods. Code is available at: https://github.com/seizetheday00/Center-Aware-DA.
CenterDA: Center-Aware Unsupervised Domain Adaptation Regularized by Class Diversity for Distracted Driver Recognition
24.09.2023
1563452 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch