Dynamic Weighting and Boundary-Aware Active Domain Adaptation for Semantic Segmentation in Autonomous Driving Environment

Semantic segmentation techniques play a crucial role in autonomous driving systems, especially in accurately parsing the environment and ensuring safe decision-making. Despite significant progress in this technology, challenges remain in terms of the cost of labeling data and the adaptability of the model when deployed to multiple target domains. Active Domain Adaptation (ADA) techniques integrate active learning and domain adaptation to adapt models from the source domain to the target domain while actively selecting the most helpful samples for labeling. However, current ADA approaches still face challenges in exploring more efficient sample selection mechanisms and mitigating data imbalance problems. This study proposes an innovative framework: Dynamic Weighting and Boundary-Aware Active Domain Adaptation (DWBA-ADA) to enhance cross-domain semantic segmentation in an autonomous driving environment. The framework emphasizes the simultaneous use of uncertainty and diversity information to identify critical image regions. The framework is improved in two ways: first, increasing the sampling proportion of minority categories, thus improving the model’s accuracy in recognizing them; and second, focusing on the identification of boundary regions across categories, thus enhancing the model’s adaptability in complex scenarios. Domain adaptation experiments across various autonomous driving datasets (e.g., GTAV, SYNTHIA, and Cityscapes) show that our DWBA-ADA approach improves the performance by 0.3% and 0.9%, respectively, compared to the current state-of-the-art MADAv2 approach under the same annotation budget and network architecture constraints. The code is available at https://github.com/licongguan/DWBA-ADA.