Semantic Segmentation using Class Aware Dimension Pooling Transformer

Semantic segmentation performance significantly deteriorates under adverse weather conditions, posing a major obstacle for autonomous vehicles that rely on accurate scene understanding. This research directly addresses the issue of poor segmentation accuracy in these challenging environments by introducing a novel architecture, the Class-Aware Dimension-Pooling Attention (CADPA) Transformer. The proposed model enhances feature representation through a multi-scale transformer encoder and a Class-Aware Context Enhancement Module, enabling better discrimination of object classes under degraded visibility. Trained on the Adverse Conditions Dataset with Correspondences (ACDC) using cross-entropy loss and the AdamW optimizer, the model is evaluated using pixel accuracy and mean Intersection over Union (mIoU). Experimental results show that CADPA achieves an mIoU of 58%. The approach effectively captures long-range dependencies and maintains spatial consistency, making it well-suited for real-world driving applications. Future work will explore self-supervised domain adaptation to improve generalization in unseen weather conditions.