OctNet is a novel 3D representation for deep learning with sparse 3D data. It enables deep and high-resolution 3D convolutional networks by hierarchically partitioning space using unbalanced octrees. Each leaf node stores a pooled feature representation, allowing efficient memory and computation allocation to dense regions. This approach reduces computational and memory requirements, enabling high-resolution learning. The paper demonstrates OctNet's utility in 3D tasks like classification, orientation estimation, and semantic segmentation. It shows that OctNet achieves higher input resolutions with lower memory consumption compared to dense networks, while maintaining performance. OctNet also provides significant speed-ups at high resolutions. The method is implemented using a hybrid grid-octree data structure, which efficiently encodes and accesses data. The paper evaluates OctNet on three tasks: 3D classification, orientation estimation, and semantic segmentation. Results show that higher resolutions improve performance, especially for orientation estimation and point cloud labeling. OctNet outperforms existing methods in terms of efficiency and performance at high resolutions. The paper also discusses related work, including dense and sparse models, and highlights the advantages of OctNet in handling sparse 3D data. The proposed method is efficient, scalable, and suitable for high-resolution 3D learning tasks.OctNet is a novel 3D representation for deep learning with sparse 3D data. It enables deep and high-resolution 3D convolutional networks by hierarchically partitioning space using unbalanced octrees. Each leaf node stores a pooled feature representation, allowing efficient memory and computation allocation to dense regions. This approach reduces computational and memory requirements, enabling high-resolution learning. The paper demonstrates OctNet's utility in 3D tasks like classification, orientation estimation, and semantic segmentation. It shows that OctNet achieves higher input resolutions with lower memory consumption compared to dense networks, while maintaining performance. OctNet also provides significant speed-ups at high resolutions. The method is implemented using a hybrid grid-octree data structure, which efficiently encodes and accesses data. The paper evaluates OctNet on three tasks: 3D classification, orientation estimation, and semantic segmentation. Results show that higher resolutions improve performance, especially for orientation estimation and point cloud labeling. OctNet outperforms existing methods in terms of efficiency and performance at high resolutions. The paper also discusses related work, including dense and sparse models, and highlights the advantages of OctNet in handling sparse 3D data. The proposed method is efficient, scalable, and suitable for high-resolution 3D learning tasks.