EPCP: An Efficient Point Cloud Classification Network with Position Fusion

Recently, deep learning on 3D point cloud data has driven rapid development in the field of autonomous driving. However, processing 3D point clouds is a challenging task. Transformers are widely used in various fields of deep learning. In the community of 3D point clouds, Point Cloud Transformer(PCT, a transformer-based framework) also achieves excellent performance on 3D point cloud tasks. However, the well-known attention mechanism benefits performance improvement while it also leads to expensive computational costs. To alleviate the problem, we propose a novel structure named efficient Point Cloud Classification Network with Position Fusion(EPCP), which uses the strategy of fusing absolute coordinates-based features and relative position information instead of the attention mechanism. By fusing relative position information, our method ensures that the output features maintain different sensitivity to specific areas, which plays a similar role as the attention mechanism. To verify the effectiveness of our method, we conduct experiments on ModelNet40 and real-world ScanObjectNN datasets. Compared with PCT, our method maintains the same accuracy on ModelNet40 and improves by 5.5% on ScanObjectNN. At the same time, our method reduces the number of parameters by 91% and FLOPs by 50% compared to PCT on the ModelNet40 dataset.