QuadFormer: Real-Time Unsupervised Power Line Segmentation with Transformer-Based Domain Adaptation

Accurately identifying Power Lines (PLs) is crucial for ensuring the safety of aerial vehicles. Despite the potential of recent deep learning approaches, obtaining high-quality ground truth annotations remains a challenging and labor-intensive task. Unsupervised Domain Adaptation (UDA) emerges as a promising solution, leveraging knowledge from labeled synthetic data to improve performance on unlabeled real images. However, existing UDA methods often suffer of huge computation costs, limiting their deployment on real-time embedded systems commonly utilized on aerial vehicles. To mitigate this problem, this paper introduces QuadFormer, a real-time framework designed for unsupervised semantic segmentation within the UDA paradigm. QuadFormer integrates a lightweight transformer-based segmentation model with a cross-attention mechanism to narrow the gap between a labelled synthetic domain and unlabelled real domain. Furthermore, we design a novel pseudo label scheme to enhance the segmentation accuracy of the unlabelled real data. To facilitate the evaluation of our framework and promote reserach in PL segemntation, we present two new datasets: AutelPL Synthetic and AutelPL Real. Experimental results demonstrate that QuadFormer achieves state-of-the-art performance on both AutelPL Synthetic → TTPLA and AutelPL Synthetic → AutelPL Real tasks. We will publicly release the dataset to the research community.