<p>In recent years, obtaining useful information from 3D point cloud to enhance the effectiveness of scene understanding tasks has become an important research topic in computer vision. To address this challenge, we propose a semantic segmentation model designed to extract distinguishing features from large-scale 3D point clouds. The model is comprised of three main modules: visual guidance generation module, co-attention fusion module, and point feature enhancement module. Firstly, the visual guidance generation module explicitly embeds color features by designing a reconstruction auxiliary network, which generates visual guidance embedding. Secondly, the co-attention module efficiently fuses these visual guidance features to top-level features of the segmentation network. Thirdly, a point feature enhancement module is applied in each decoding layer, based on the active accumulation of the neighborhood features, improving the segmentation of boundary points. This proposed module can be seamlessly integrated into various 3D point cloud understanding architectures. Extensive experiments have been conducted on two challenging datasets, and the results demonstrate that our method outperforms several state-of-the-art methods in the majority of cases.</p>