Predicts per-voxel class labels for a scan, which are labeled column-by-column. Code is adapted from the semantic voxel labeling task of ScanNet, but uses only occupancy information in a 31x31x62 neighborhood to labeling the 1x1x62 center column.
Training uses Torch7, with torch packages cudnn, cunn, hdf5, xlua.
th train.lua --train_data [path to train h5 file list] --test_data [path to test h5 file list] --save [output path]
with the appropriate paths to the train/test data (use '--help' to see more options)