Multitask Adaptation for Unlabeled Domain Using Multiple Single-Task Domains

Semantic segmentation and depth estimation tasks are crucial for autonomous driving systems, but obtaining their labels from real-world datasets is costly. To address the problem, we developed a multitask domain adaptation that uses various labeled datasets with distinct tasks to adapt the multitask model for the unlabeled domain. The proposed framework can handle multiple source domains containing various task labels, which allows us to extend the combinations of acceptable source datasets in contrast to the previous multitask domain adaptation methods. We suggest using the ‘TripleMix’ approach to obtain the integrated features from the three separate domains, including two labeled domains and one unlabeled domain. In addition, we design a task correlation network that trains multiple tasks through attentional correlation, increasing the synergies between various tasks. To validate the proposed algorithm’s state-of-the-art performance based on the interactions of the different domains and tasks, we analyze it using a variety of dataset combinations that consider two virtual domains and one real-world target domain.