Skip to content

Training PN++ #35

Description

@saidharb

Hi Chris,

I really like your work and currently I am trying to recreate your metrics for CAD-model reconstruction from point clouds. I built my own training script for PN++, however I cannot achieve the metrics indicated in the supplementary document in your paper.

After scrutinizing pc2cad_train.py I saw that you train on point clouds with normals, however in the pc2cad.py script no normals are used. Therefore my 3 questions:

  • Was pc2cad_train.py or pc2cad.py used for PN++ training?
  • Did you use normals to train PN++? If so how were the normals obtained, because the script json2pc.py only creates point clouds with cartesian coordinates and no normals.
  • What was your final validation MSE loss after 200 epochs of training? I reached approximately 0.0705, but my ACC_cmd is only at 74.82 (yours at 85.95) and ACC_param at 66.80 (yours 74.22). The provided pre-trained DeepCAD model cannot be the bottleneck, as I tested it and was able to recreate the metrics in the paper.

Thank you in advance for your effort.

Best regards
Said

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions