Efficient 3D point cloud network model for reconstruction of 3D point cloud, object classification, and part segmentation

This study proposes three efficient 3D point cloud network models for reconstruction, object classification, and part segmentation to reduce costs and labor in 3D modeling. The proposed models achieve high accuracy and performance by optimizing 3D-LMNet with improved encoder depth, point attention mechanisms, and refined MLP structures. Extensive experiments on G3D and S3D datasets using 1024 and 2048 points demonstrate the effectiveness of the architectures. The best-performing configurations for encoders and decoders are identified across tasks. The results show that the proposed models significantly enhance modeling efficiency while maintaining high precision, making them suitable for real-world 3D applications.