Volumetric and Multi-View CNNs for Object Classification on 3D Data Supplementary Material

Training for Our Volumetric CNNs To produce occupancy grids from meshes, the faces of a mesh are subdivided until the length of the longest edge is within a single voxel; then all voxels that intersect with a face are marked as occupied. For 3D resolution 10,30 and 60 we generate voxelizations with central regions 10, 24, 54 and padding 0, 3, 3 respectively. This voxelization is followed by a hole filling step that fills the holes inside the models as occupied voxels. To augment our training data with azimuth and elevation rotations, we generate 60 voxelizations for each model, with azimuth uniformly sampled from [0, 360] and elevation uniformly sampled from [−45, 45] (both in degrees). We use a Nesterov solver with learning rate 0.005 and weight decay 0.0005 for training. It takes around 6 hours to train on a K40 using Caffe [2] for the subvolume supervision CNN and 20 hours for the anisotropic probing CNN. For multi-orientation versions of them, SubvolumeSup splits at the last conv layer and AniProbing splits at the second last conv layer. Volumetric CNNs trained on single orientation inputs are then used to initialize their multiorientation version for fine tuning. During testing time, 20 orientations of a CAD model occupancy grid (equally distributed azimuth and uniformly sampled elevation from [−45, 45]) are input to MO-VCNN to make a class prediction.