This paper presents two novel volumetric CNN architectures and improvements to multi-view CNNs for 3D object classification. The authors analyze the performance gap between volumetric and multi-view CNNs and propose solutions to bridge this gap. They introduce two volumetric CNNs: one with auxiliary tasks to improve learning of local features, and another using anisotropic kernels to capture long-range interactions. They also enhance multi-view CNNs with multi-resolution filtering and data augmentation. The results show that their volumetric CNNs can match the performance of multi-view CNNs at a 30x30x30 resolution. The authors also introduce a new real-world 3D dataset for evaluation and demonstrate that their methods outperform previous approaches on both synthetic and real-world data. The study highlights the importance of 3D resolution and network architecture in achieving high performance in 3D object classification.This paper presents two novel volumetric CNN architectures and improvements to multi-view CNNs for 3D object classification. The authors analyze the performance gap between volumetric and multi-view CNNs and propose solutions to bridge this gap. They introduce two volumetric CNNs: one with auxiliary tasks to improve learning of local features, and another using anisotropic kernels to capture long-range interactions. They also enhance multi-view CNNs with multi-resolution filtering and data augmentation. The results show that their volumetric CNNs can match the performance of multi-view CNNs at a 30x30x30 resolution. The authors also introduce a new real-world 3D dataset for evaluation and demonstrate that their methods outperform previous approaches on both synthetic and real-world data. The study highlights the importance of 3D resolution and network architecture in achieving high performance in 3D object classification.