This paper addresses the task of object classification on 3D data using volumetric and multi-view Convolutional Neural Networks (CNNs). The authors analyze the performance gap between volumetric and multi-view CNNs, which is attributed to differences in network architecture and 3D resolution. They propose two new architectures for volumetric CNNs: one introduces auxiliary training tasks to focus on partial subvolumes, and the other uses anisotropic kernels to capture long-distance interactions. Additionally, they introduce multi-resolution filtering in multi-view CNNs to improve performance. Extensive experiments on the ModelNet dataset and a new real-world reconstruction dataset demonstrate that their proposed methods outperform state-of-the-art approaches, closing the gap between volumetric and multi-view CNNs. The paper also explores the impact of 3D resolution and provides insights into future directions for improving volumetric CNNs.This paper addresses the task of object classification on 3D data using volumetric and multi-view Convolutional Neural Networks (CNNs). The authors analyze the performance gap between volumetric and multi-view CNNs, which is attributed to differences in network architecture and 3D resolution. They propose two new architectures for volumetric CNNs: one introduces auxiliary training tasks to focus on partial subvolumes, and the other uses anisotropic kernels to capture long-distance interactions. Additionally, they introduce multi-resolution filtering in multi-view CNNs to improve performance. Extensive experiments on the ModelNet dataset and a new real-world reconstruction dataset demonstrate that their proposed methods outperform state-of-the-art approaches, closing the gap between volumetric and multi-view CNNs. The paper also explores the impact of 3D resolution and provides insights into future directions for improving volumetric CNNs.