CA-Net: a context-awareness and cross-channel attention-based network for point cloud understanding

To capture local complex shape information and fine-grained features from irregular point cloud, this paper proposes a novel local feature encoder-based network named context-awareness (CA)-Net which is used to solve the challenge of 3D object classification and segmentation. The core of the CA-Net is CA-Encoder, which is based on CA and cross-channel multi-head self-attention (CC-MSA). CA-Encoder uses contextual information awareness to aggregate local features from two levels: point cloud 3D coordinate information and high-dimensional implicitly encoded information, leveraging CC-MSA to learn channel-related information. For different point cloud benchmark tasks, CA-Net uses the local feature enhancement module (classification) and the Up-Transformer (segmentation) which includes a cascaded set of CA-Encoders to solve the problem of feature loss at non-edge points in edge sampling, so that the sampled results can both preserve the shape of the point cloud edges and reconstruct the full internal shape structure of the point cloud. The CA-Net has superior performance in experiments on ModelNet and ShapeNetPart datasets with a classification accuracy of 93.8% and a segmentation accuracy of 85.9%.