SIGNATURE WORK
CONFERENCE & EXHIBITION 2024

Semantic segmentation plays a pivotal role in computer vision, especially for critical applications like autonomous driving and medical imaging. To overcome the challenges of data annotation bottlenecks in semantic segmentation, semi-supervised learning leverages both labeled and unlabeled data, with the quality of pseudo-labels being crucial to model performance. Our study introduces a novel network, Uncertainty-aware Unbiased Subclass Regularization (UA-USRN), an integration of the Latent Discriminant Deterministic Uncertainty (LDU) method with the Unbiased Subclass Regularization Network (USRN), specifically designed to quantify the uncertainty and enhance the reliability of pseudo-label quality. Synergizing the strengths of LDU and USRN, our work not only boosts the accuracy and reliability of pseudo-labels but also effectively tackles the class imbalance issue, leading to significant performance improvements. Rigorous experimentation on the PASCAL VOC benchmark demonstrates UA-USRN’s substantial superiority over both the semi-supervised USRN model and supervised baselines, both qualitatively and quantitatively. This underscores our approach’s real-world applicability and establishes new benchmarks for accuracy and reliability in semantic segmentation.