We recommend you to use anaconda to make sure that all dependencies are in place. The code is tested in the following setting:
- python == 3.8
- pytorch == 1.12.0
- CUDA == 11.3
- Packages: trimesh == 3.12.7, omegaconf == 2.2.2, tqdm == 4.64.0, scikit-image == 0.19.3, plotly == 5.9.0
The ABC dataset is used in this paer, you can visit their official website to download it.
We also provide several examples for testing our network.
MixNet can be used to reconstruct a high-quality surface given a point cloud with normal data. Adjust config/config.yaml to the path of the input point cloud:
dataio:
...
data_path: your data path
...
Then run the training script:
python runner.py
You can change the GPU indexes and network configurations depending on your own needs:
python runner.py --gpu INDEX --conf_path YOUR_OWN_CONF
According to config/config.yaml:
dataio:
expname: recon
exps_folder_name: exp
...
After finishing the training, the reconstrution results can be found in the export folder: ..\exp\recon\.
