About the ScanNetPP preprocessing

[RayYoh]

Hi authors,

I find that the segments for ScanNet++ val set are different from the official toolkit results. This is an example.

[Xiaoyang Wu]

Hi, good point. Could you provide the configuration used to create processed datasets with the official toolkit? (I wrote the preprocessing code based on my personal understanding towards this dataset.)

[RayYoh]

Hi Xiaoyang, the yaml file I used for preprocessing is:

data:
  data_root: '/path/to/scannetpp/data'

  labels_path: '/path/to/scannetpp/metadata/semantic_benchmark/top100.txt'
  # for instance segmentation
  use_instances: true
  instance_labels_path: '/path/to/scannetpp/metadata/semantic_benchmark/top100_instance.txt'

  ## save multiple labels per vertex/point? ##
  # multilabel:
  #   max_gt: 3
  #   multilabel_only: false

  mapping_file: '/path/to/scannetpp/metadata/semantic_benchmark/map_benchmark.csv'

  list_path: '/path/to/scannetpp/splits/nvs_sem_val.txt'
  # list_path: '/mnt/nvme1/ray/dataset/scannetpp/splits/sem_test.txt'

  ignore_label: -1

  sample_factor: 1.0

  transforms:
    # read the mesh 
    - add_mesh_vertices
    # map raw labels to benchmark classes
    - map_label_to_index
    # use segments info to get labels on the vertices, handle multilabels
    - get_labels_on_vertices
    - add_normals
    # sample points on the mesh and transfer all vertex info to the points
    # - sample_points_on_mesh

# dir to save pth training data
out_dir: '/path/to/scannetpp_pth/val_vtx'

And then I use the official toolkit to run:
python -m semantic.prep.prepare_training_data semantic/configs/prepare_training_data.yml.
Note that I comment sample_points_on_mesh. Because I find that you provided code seems no this operation (is it correct?).

The coord and color (val_vtx color * 255) are same.