DANF: Dual-Flow Aggregation Network for Curtain Wall Frame Real-Time Segmentation

The precise positioning of curtain wall frames is a crucial step for the automated installation of curtain walls. This paper proposes a network (DANF) with a dual-flow aggregation structure to achieve semantic segmentation of curtain wall frames. We constructed a dataset (CWF) for the segmentation of curtain wall frames and analyzed the characteristics of both the curtain wall frames and the background environment. Our proposed network architecture combines the strengths of transformers for semantic information extraction and CNNs for high-resolution feature extraction. Moreover, we introduced a dual-flow aggregation module to effectively fuse the features derived from transformers and CNNs. The experimental results on the CWF dataset validate the powerful performance of our method.