Deep Learning Architectures for 3D Shape Classification: A Study on PointNet and Voxel Grid-Based Models

Three-Dimensional (3D) models are pivotal in various domains, including computer vision, robotics, medical imaging, and virtual reality, offering a foundation for diverse applications in the 3D world. Representing 3D data involves several formats, such as point clouds, voxel grids, polygonal meshes, and depth images, each characterized by unique computational and structural properties. Unlike the structured nature of 2D image grids, 3D data often exhibits irregular and unordered structures, presenting significant challenges for deep learning models. This study provides a comprehensive comparative analysis of state-of-the-art neural network architectures tailored for processing 3D data, focusing on their effectiveness in 3D model classification. Using the widely recognized ModelNet10 dataset, we investigate the impact of different 3D data representations, specifically point clouds and voxel grids, on 3D model classification performance. We implement the PointNet architecture for point cloud representation and evaluate voxel grid-based architectures, including ModelNet-Cnn, VoxNet, and Voxception-ResNet42. Our findings reveal that PointNet achieves superior classification results, attaining a validation accuracy of 90%, outperforming voxel grid-based models such as ModelNet-Cnn (87%), VoxNet (79%), and Voxception-ResNet42 (76%). These results underscore the advantages of point cloud representations for 3D shape classification and highlight the critical role of representation choice in optimizing model performance.