Orientation-Aware Multi-Scale Fusion Algorithm for Point Cloud Classification

Recently, in the field of point cloud processing, a two-stage strategy of applying pre-training followed by fine-tuning for various downstream tasks has achieved inspiring performance in tasks such as object classification and semantic segmentation. However, significant challenges remain in terms of fine-grained feature capture efficiency and representation capability in sparse regions. To address these issues, this paper proposes an orientation-aware multi-scale fusion algorithm for point cloud classification. In terms of feature extraction, we utilize orientation-aware Octant-constrained K-Nearest Neighbors (O-KNN) and enhance the feature representation of sparse regions by concatenating multi-scale point cloud patches. During pre-training and fine-tuning phases, we learn point cloud contextual representations through the Masked Point Modeling (MPM) task. Furthermore, to address feature shift and information redundancy that may arise during feature extraction, we propose a key-value subspace adaptive selection mechanism, combined with root mean square (RMS) normalization, to mitigate the impact of these issues. The overall training process utilizes a Teacher-Student architecture and Exponential Moving Average (EMA) for self-distillation learning, ensuring robust and stable results. Experimental results demonstrate that our method achieves promising performance on the ModelNet40 classification task and the ShapeNetPart part segmentation task, confirming the dual advantages of our approach in feature extraction and model lightweighting.