Edge-aware distilled segmentation with pseudo-label refinement for autonomous driving perception

Abstract

Achieving precise semantic segmentation is essential for enabling real-time perception in autonomous systems, yet leading approaches typically require substantial annotated data and powerful hardware, restricting their use on devices with limited resources. This work introduces an efficient segmentation framework that integrates pseudo-label refinement, knowledge distillation, and entropy-based confidence filtering to train compact student networks suitable for edge deployment. High-quality pseudo-labels are first produced by a robust teacher network, then further improved using a dense conditional random field to boost spatial consistency. An entropy-based selection mechanism removes unreliable predictions, ensuring that only the most trustworthy labels guide the student model's training. The use of knowledge distillation effectively transfers detailed semantic understanding from the teacher to the student, enhancing accuracy without added computational overhead. Experimental results with multiple EfficientNet backbones reveal that this pipeline improves segmentation accuracy and output clarity, while also supporting real-time or near real-time inference on CPUs with limited processing power. Extensive ablation and qualitative studies further confirm the method's robustness and flexibility for real-world edge applications.