OctNet is a novel representation for deep learning with sparse 3D data, enabling 3D convolutional networks that are both deep and high-resolution. Unlike existing models, OctNet hierarchically partitions 3D space using unbalanced octrees, where each leaf node stores a pooled feature representation. This approach focuses memory allocation and computation on relevant dense regions, allowing deeper networks without compromising resolution. The authors demonstrate the utility of OctNet by analyzing its impact on various 3D tasks, including 3D object classification, orientation estimation, and point cloud labeling. They show that OctNet enables significantly higher input resolutions compared to dense inputs due to lower memory consumption, achieving identical performance at lower resolutions while providing speed-ups at resolutions of $128^3$ and above. The code for OctNet is available on GitHub.OctNet is a novel representation for deep learning with sparse 3D data, enabling 3D convolutional networks that are both deep and high-resolution. Unlike existing models, OctNet hierarchically partitions 3D space using unbalanced octrees, where each leaf node stores a pooled feature representation. This approach focuses memory allocation and computation on relevant dense regions, allowing deeper networks without compromising resolution. The authors demonstrate the utility of OctNet by analyzing its impact on various 3D tasks, including 3D object classification, orientation estimation, and point cloud labeling. They show that OctNet enables significantly higher input resolutions compared to dense inputs due to lower memory consumption, achieving identical performance at lower resolutions while providing speed-ups at resolutions of $128^3$ and above. The code for OctNet is available on GitHub.