Abstract: Aiming at the problems of blur and distortion of glass curtain wall in urban 3D model, this paper proposes a 3D glass curtain wall segmentation method based on projection and feature activation. The method firstly extracts building facade information by layers using point cloud feature. Then the glass facade is projected onto the aerial image using a projection based segmentation method. Finally, the feature activation module is nested on U-Net network structure to achieve accurate extraction of glass curtain wall. By comparing the proposed method with U-Net and U²-Net models, the new method proposed in this paper can not only extract clear glass region boundaries, but also effectively reduce the influence of oblique aerial images on small glass regions and improve segmentation performance, which is of great significance for realizing low-cost fine modeling.

Key words: 3D real scene, glass curtain wall extraction, partition, convolutional neural network