Lin Y.-PYeh Y.-MChou Y.-CYI-CHANG LU2023-06-092023-06-092021https://www.scopus.com/inward/record.uri?eid=2-s2.0-85126663261&partnerID=40&md5=0e5f5f7cbf287304462b01bf80531620https://scholars.lib.ntu.edu.tw/handle/123456789/632092Due to the irregular and unordered properties of 3D point cloud data, it is more challenging to extract geometric features between points. In this paper, we propose techniques to improve the ability of capturing point cloud features, so that higher accuracy and better stability of point cloud classification tasks can be achieved. First, we design two attention modules: the Point-wise Attention Module determines the correlation between points, and the Channel-wise Attention Module allows the model to focus on important features with limited resources. With these attention modules, we not only achieve the state-of-the-art accuracy of 93.7% on the ModelNet40 dataset but also reduce the error rates on the ScanObjectNN dataset. Secondly, we propose a guideline to dynamically adjust the size of the KNN. Using the proposed Dynamic-K method, we can significantly increase the accuracy of classification when dealing with low-resolution objects. © 2021 APSIPA.3D Point Cloud Classification; Attention Mechanism; Deep Neural NetworkClassification (of information); 3D point cloud; 3d point cloud classification; Attention mechanisms; Classification tasks; Cloud classification; Geometric feature; High-accuracy; Point cloud data; Point-clouds; Property; Deep neural networksconference paper2-s2.0-85126663261