PointGTA: A Geometric Triangle-Aware Network for 3D Point Cloud Learning

The analysis of point clouds is challenging due to the irregular and unordered nature of 3D data. Drawing inspiration from the success of geometric triangle-aware methods in protein structure prediction, we investigate the potential of geometric triangle-aware mechanisms for improving 3D point cloud understanding. To enhance local feature representation, we propose the point triangle-aware (PTA) module, which integrates the point triangular update layer and the point triangular self-attention layer. These components model the geometric relationships among neighboring points through triangular geometric relationships, thereby capturing local shape characteristics with improved precision. Additionally, to address the limitations of existing methods in capturing long-range dependencies, we introduce the long-range aware transformer (LAT) module, which extends the receptive field to incorporate broader contextual information. By combining the PTA and LAT, we present the point geometric triangle-aware network (PointGTA) for 3D point tasks such as semantic segmentation, object part segmentation, and shape classification. PointGTA effectively combines local geometric detail extraction with long-range contextual modeling, leading to robust feature learning. Experiments on multiple benchmark datasets for 3D point clouds demonstrate the effectiveness of our proposed method. Our code is publicly available at: https://github.com/maoerlu/PointGTA.