Kumari, SeemaSeemaKumariKalpesh Patel, SamaySamayKalpesh PatelMuthalagu, RajaRajaMuthalaguRaman, ShanmuganathanShanmuganathanRaman2025-08-312025-08-312025-01-01[9783031781124]10.1007/978-3-031-78113-1_292-s2.0-85212278393https://d8.irins.org/handle/IITG2025/28587Deep learning-based approaches have shown great achievement in 3D point analysis. Due to the irregular and unordered data structure, point cloud analysis is still very challenging. Most existing work uses the convolution, graph, or attention mechanism to achieve the 3D geometry of the target shape. Only a few approaches consider global and local geometry information of point clouds. However, both kinds of geometry play a significant role in analysis. This paper proposes a geometry-acquainted fusion (GAF) module that considers global-to-local geometry information by multi-step processing. Further, we consider in-plane and out-plane distances to capture the geometrical information in the raw point cloud. The modules are utilized in two different architectures, devised for classification and segmentation. The classification network is a simple feed-forward architecture, whereas the segmentation network is developed based on a U-Net-like architecture with residual connections. We show that the proposed architectures perform quite well compared to the state-of-the-art methods in classification and segmentation tasks.false3D point cloud | classification | fusion | geometry | segmentationConference Paper16113349443-45820250cpBook Series0