This is the official implementation of the paper Graph-to-3d: End-to-End Generation and Manipulation of 3D Scenes Using Scene Graphs | arxiv
Helisa Dhamo*, Fabian Manhardt*, Nassir Navab, Federico Tombari
ICCV 2021

We address the novel problem of fully-learned 3D scene generation and manipulation from scene graphs, in which a user can specify in the nodes or edges of a semantic graph what they wish to see in the 3D scene.

If you find this code useful in your research, please cite

@inproceedings{graph2scene2021,
  title={Graph-to-3D: End-to-End Generation and Manipulation of 3D Scenes using Scene Graphs},
  author={Dhamo, Helisa and Manhardt, Fabian and Navab, Nassir and Tombari, Federico},
  booktitle={IEEE International Conference on Computer Vision (ICCV)},
  year={2021}
}

We have tested it on Ubuntu 16.04 with Python 3.7 and PyTorch 1.2.0

# clone this repository and move there
git clone https://github.com/he-dhamo/graphto3d.git
cd graphto3d
# create a conda environment and install the requirments
conda create --name g2s_env python=3.7 --file requirements.txt 
conda activate g2s_env          # activate virtual environment
# install pytorch and cuda version as tested in our work
conda install pytorch==1.2.0 cudatoolkit=10.0 -c pytorch
# more pip installations
pip install tensorboardx graphviz plyfile open3d==0.9.0.0 open3d-python==0.7.0.0 
# Set python path to current project
export PYTHONPATH="$PWD"

To evaluate shape diversity, you will need to setup the Chamfer distance. Download the extension folder from the AtlasNetv2 repo and install it following their instructions:

cd ./extension
python setup.py install

Updated The checkpoints for our trained models and the Atlasnet weights (used to obtain shape features) can be found here. Once downloaded place them in the ./experiments folder and unzip.

Download the 3RScan dataset from their official site. You will need to download the following files using their script:

python download.py -o /path/to/3RScan/ --type semseg.v2.json
python download.py -o /path/to/3RScan/ --type labels.instances.annotated.v2.ply

Additionally, download the metadata for 3RScan:

cd ./GT
chmod +x ./download_metadata_3rscan.sh && ./download_metadata_3rscan.sh

Download the 3DSSG data files to the ./GT folder:

chmod +x ./download_3dssg.sh && ./download_3dssg.sh

Update: Next, to fix a few dataset changes or incompatibilities - Find and delete the following line from train_scans.txt: fa79392f-7766-2d5c-869a-f5d6cfb62fc6. This scan contains an instance with zero points (depending on your 3RScan dataset version) which can lead to a crash during training. Additionally, add _scene_ as the first line of the file classes.txt.

We use the scene splits with up to 9 objects per scene from the 3DSSG paper. The relationships here are preprocessed to avoid the two-sided annotation for spatial relationships, as these can lead to paradoxes in the manipulation task. Finally, you will need our directed aligned 3D bounding boxes introduced in our project page. The following scripts downloads these data.

chmod +x ./download_postproc_3dssg.sh && ./download_postproc_3dssg.sh

Run the transform_ply.py script from this repo to obtain 3RScan scans in the correct alignment:

cd ..
python scripts/transform_ply.py --data_path /path/to/3RScan

To train our main model with shared shape and layout embedding run:

python scripts/train_vaegan.py --network_type shared --exp ./experiments/shared_model --dataset_3RScan ../3RScan_v2/data/ --path2atlas ./experiments/atlasnet/model_70.pth --residual True

To run the variant with separate (disentangled) layout and shape features:

python scripts/train_vaegan.py --network_type dis --exp ./experiments/separate_baseline --dataset_3RScan ../3RScan_v2/data/ --path2atlas ./experiments/atlasnet/model_70.pth --residual True

For the 3D-SLN baseline run:

python scripts/train_vaegan.py --network_type sln --exp ./experiments/sln_baseline --dataset_3RScan ../3RScan_v2/data/ --path2atlas ./experiments/atlasnet/model_70.pth --residual False --with_manipulator False --with_changes False --weight_D_box 0 --with_shape_disc False

One relevant parameter is --with_feats. If set to true, this tries to read shape features directly instead of reading point clouds and feading them in AtlasNet to obtain the feature. If features are not yet to be found, it generates them during the first epoch, and reads these stored features instead of points in the next epochs. This saves a lot of time at training.

Each training experiment generates an args.json configuration file that can be used to read the right parameters during evaluation.

To evaluate the models run