Reviews: Point-Voxel CNN for Efficient 3D Deep Learning

Many improvements over previous results are numerically quite small and it is unclear if these improvements are statistically significant. This includes the following points: - No baseline result (e.g. Why not show voxel and point cloud based methods in both graphs? It is hence not clear how much of the improvement is due to the concrete implementation vs. the proposed method. Maybe it would even make sense to combine the proposed method with sparse convolutions?