DATA.md

Prepare Data

Download Dataset

🏠 Scene: We train and evaluate the scene layout based on 3D FRONT dataset.

🪑 Rigid Object: We then utilize 3D FUTURE as the rigid objects in the generated scene.

Download 3D FRONT and 3D FUTURE dataset with the instruction in this website.

💻 Articulated object: We utilize articulated objects in GAPartNet dataset for the generated scene.

Download GAPartNet dataset following the instructions in this website.

Preprocess data

Pickle scene & object dataset

sh run/pickle_3dfuture_dataset.sh save_dir 
# for example:
# sh run/pickle_3dfuture_dataset.sh data/pickled_data  

You will get (1) pickled scene dataset in PATH_TO_SCENES defined in the config file, and (2) pickled object dataset in save_dir/threed_future_model_roomtype.pkl.

Pickle mesh info of each CAD model in both 3D FUTURE and GAPartNet

sh run/pickle_pcd.sh save_dir
# for example:
# sh run/pickle_pcd.sh data/pickled_data

You will get preprocessed pointcloud, saved as .npz and .ply, for each object in 3D FUTURE and GAPartNet.

Train Autoencoder, and generate latent feature

sh run/objautoencoder.sh save_dir experiment_tag 
# for example
# sh run/objautoencoder.sh autoencoder_output debug

You will get geometric latent feature, saved as raw_model_norm_pc_lat32.npz for each object.

Preprocess scene dataset of 3D FRONT

sh run/preprocess_data.sh save_dir
# for example:
# sh run/preprocess_data.sh data/preprocessed_data

You will get room info and rendered image in save_dir. Here we set background color as gray for better visualization.