downsampling

Downsampling

Uniformly downsampling of the point cloud result in more representative blocks because regions within the block of high density are thinned out whereby the effect to regions of lower density is not that significant.

Training and validation is done using these downsampled point clouds.

For the final evaluation knn interpolation is used to retrieve class labels for all points of the full sized point cloud.

To increase the training speed, the downsampled point cloud are pre-calculated and stored within the directory structure (see dedicated chapter of wiki):

python tools/downsample.py --data_dir path/to/dataset --cell_size 0.01