search

CuiRuikai / Partial2Complete

[ICCV 2023] P2C: Self-Supervised Point Cloud Completion from Single Partial Clouds
MIT License
159 stars 9 forks source link

How to create a self supervised dataset. How should I train without using a complete point cloud as the ground truth. #20

Open 3058356235 opened 6 months ago

3058356235 commented 6 months ago

I created my own dataset based on the EPN3D format, and I noticed that your paper does not require a complete point cloud as a ground truth. However, in the dataset in this format, the complete folder must contain the complete ground truth, otherwise the network will not be able to train correctly and will prompt that the num_samples used for training is 0.

CuiRuikai commented 6 months ago

The complete folder is used for evaluation, so you can simply set it as the same as partial path. That is, for cfgs/dataset_configs/EPN3D.yaml, set

COMPLETE_POINTS_PATH: ./data/EPN3D/%s/complete/%s.npy

as 

COMPLETE_POINTS_PATH: ./data/EPN3D/%s/partial/%s.npy
3058356235 commented 6 months ago

Cheers for you, it's too responsible. Special and timely response, thank you again.