Abstract:Due to the sparseness and irregularity of the point cloud, accurate extraction of internal structural details from the point cloud as well as fast identification of the overall contour remains a challenging task. Currently, most studies focus on introducing sophisticated designs to unilaterally capture local or global features of point cloud, and rarely combine local features with global features. More importantly, it is easy to increase the computational burden while pursuing efficiency. We propose a lightweight feature extractor that efficiently extract and fuse local and global features of point cloud, which is named as AFpoint. Specifically, AFpoint is composed of two modules: the Local-Global Parallelized Feature Extraction module (LGP) and the Adaptive Feature Fusion module (AFF). The LGP module encodes local and global features in parallel by using point-by-point convolution and relative attention mechanism, respectively. It simultaneously performs the task that extracts the fine-grained structure and captures the global relationships. The AFF module adaptively selects and integrates the local and global features by estimating the attention maps of encoded features and helps the model to autonomously focus on important regions. Note that AFpoint is a plug-and-play and universal module. We use AFpoint to construct classification and segmentation networks of point cloud, which greatly improves the accuracy and robustness of the baseline model and reduces the parameters by nearly half. Experiments on the widely adopted ModelNet40, ScanObjectNN classification dataset demonstrate the state-of-the-art performance of our network and also show good results on experiments with the ShapeNetPart part segmentation dataset.

AFpoint: adaptively fusing local and global features for point cloud