Comprehensive Teacher-based Multi-task Model Train Using a Small Amount of Data

Yuchang Zhu,Nanfeng Xiao

DOI: https://doi.org/10.1109/ijcnn60899.2024.10651103

2024-01-01

Abstract:In the autonomous driving, effectively managing multiple visual perception tasks simultaneously is essential. Multi-task processing models are thus crucial in this field. The primary challenge lies in training these models, which typically require extensive and meticulously labeled datasets, a process of-ten constrained by resources. This challenge is particularly acute with multi-task datasets, where accurate labeling of numerous data samples becomes a time-intensive endeavor. To overcome this, we propose Comprehensive Teacher, a novel semi-supervised method specifically designed for multi-task model training. This method innovatively leverages the principles of pseudo-label learning and consistency regularization, traditionally used in semi-supervised learning, and adapts them to multi-task contexts. Our approach demonstrates significant versatility and efficiency in the training multi-task models. We validate Comprehensive Teacher through extensive experiments on diverse autonomous driving datasets, showcasing its superiority over conventional single-task semi-supervised methods. This approach enables the training of high-quality multi-task models with limited data availability, marking a significant step forward in resource-efficient autonomous driving technologies.