Point Cloud Classification Based on Local Fusion and Dense Connection Structure

Due to the influence of objective factors such as scene noise and hardware equipment, compared to CAD modeling point clouds, the real point clouds collected by LiDAR often have strong non-uniformity and irregularity, which leads to the low ability of traditional point cloud classification models to extract point cloud features in real scenes and poor classification robustness. To address this issue, we have added a local feature fusion module and a dense connected feature fusion method to PointMLP, thereby encoding broader point cloud information into local features. The simple stacking of multiple MLP and Mixer blocks constitutes our network model-PointMM. The shallow features of the point cloud are used as input and enter the local feature fusion module to perform feature fusion within and between point sets. Finally, by maximizing the pooling layer and selecting points with stronger semantic features as inputs to the second stage feature extraction module. This method has been tested on the ScanObjectNN dataset and shows that it can approach SOTA with a classification accuracy of 0.8%, surpassing other mainstream models.