Exploring multi-scale and cross-type features in 3D point cloud learning with CCMNET

Abstract

The existing methods for 3D point cloud learning can be broadly categorized into point-based and voxel-based approaches. Typically, these techniques often produce features that are either overly fine-grained or excessively coarse-grained. Moreover, most conventional methods primarily concentrate on extracting multi-scale information from a single feature type, overlooking the potential advantages of integrating diverse multi-scale features. To overcome these limitations, we propose CCMNet, an innovative framework for 3D point cloud learning that leverages Coarse-to-fine and Cross-type Multi-scale features. CCMNet integrates three levels of feature granularity: coarse-grained, mid-grained, and fine-grained. Coarse-grained features are extracted using a 3D CNN with low voxel resolution, mid-grained features are captured through an attention mechanism operating both within and across neighborhoods, and fine-grained features are derived using a streamlined multi-layer perceptron (MLP) network. In addition, we introduce a cross-type multi-scale strategy to enhance local feature representations by seamlessly integrating features across different scales and types. CCMNet serves as the feature extraction network for point cloud classification and segmentation tasks. Experimental results highlight that our method achieves significant performance improvements in 3D point cloud learning. The source code is publicly available at https://github.com/NWUzhouwei/CCMNet.