NAP: Neural Articulated Object Prior

project page Neurips-2023

[2023.Oct.21] Note: because we are a little busy recently, we are still working on the full releaseing, currently the repo is a preview, supporting basic training and inference

Install

Run bash env.sh, this will create a conda environment named nap-gcc9 and install all dependencies. This script is tested with Ubuntu 20.04 and cuda 11.7.

Prepare data and checkpoints

Currently, we only release the pre-processed training data for articulated objects. The part shape prior data is not released yet.

Download pre-processed data from link. Unzip it and form the directory structure as follows:

PROJECTROOT/data
├── partnet_mobility_graph_mesh
└── partnet_mobility_graph_v4

Download the pretrained checkpoint (necessary for training NAP because it contains the shape prior network weights) from link. Put them under PROJECTROOT/log/:
```
PROJECTROOT/log
├── s1.5_partshape_ae
└── v6.1_diffusion
```
Optionally, you can download the evaluation output example from link. Unzip it and put it under PROJECTROOT/log/test/:
```
PROJECTROOT/log/test
├── G
├── ID_D_matrix
├── PCL
└── Viz
```

Again, here are the downloading links

training-data-download

checkpoint-download

test-output-example-download

Training

One training example is:

python run.py --config ./configs/nap/v6.1_diffusion.yaml -f

You can also check the .vscode/launch.json.

Testing

Computing the metrics takes some time, please see eval/readme_eval.md for details.

Note

If you find this repo useful, please cite our paper. Thank you!