This repo implements a BSP-CVAE model, which uses the idea of BSP-Net but is a generative model.

• reconstruction on ScanNet:

• generation (interpolated latent codes):

This repo is tested on Ubuntu16.04, CUDA 10.1.
For the python dependencies, see requirements.txt.
We also use two Cython extensions, install them by python setup.py build_ext --inplace.

conda env create -n bsp
conda activate bsp
pip install -r requirements.txt
python setup.py build_ext --inplace

Data are assumed to be located in ${data_root}/datasets/, where ${data_root} can be set in main.py.

• ShapeNet
  We use the preprocessed data provided by RfDNet, please follow their instructions and put it under ${data_root}/datasets/ShapeNetv2_data. We use 8 classes ('table', 'chair', 'bookshelf', 'sofa', 'trash_bin', 'cabinet', 'display', 'bathtub') in ShapeNet to train the model.

• ScanNet & Scan2CAD
  If you want to test the reconstruction performance under indoor scenes, you should also download preprocessed ScanNet and Scan2CAD datasets following these instructions, and put it under ${data_root}/datasets/scannet.

The options and parameters should be modified directly in main.py.

# train with default settings.
bash train.sh

It takes about 4 days to train the model for 800 epochs on a single GPU.

Logs are saved in workspace/log_${exp_name}.txt.
Checkpoints are saved in workspace/checkpoints.
Tensorboard records are saved in workspace/run.

After training, just run the tests as follows:

# reconstruction on shapenet
CUDA_VISIBLE_DEVICES=1 OMP_NUM_THREADS=1 python main.py --checkpoint latest --test_shapenet
# generation
CUDA_VISIBLE_DEVICES=1 OMP_NUM_THREADS=1 python main.py --checkpoint latest --generate
# interpolated generation
CUDA_VISIBLE_DEVICES=1 OMP_NUM_THREADS=1 python main.py --checkpoint latest --interpolated_generate
# reconstruction on scannet
CUDA_VISIBLE_DEVICES=1 OMP_NUM_THREADS=1 python main.py --checkpoint latest --test_scannet
# save zs
CUDA_VISIBLE_DEVICES=1 OMP_NUM_THREADS=1 python main.py --checkpoint latest --save_z
# save database
CUDA_VISIBLE_DEVICES=1 OMP_NUM_THREADS=1 python main.py --checkpoint latest --save_db

Results are saved in workspace/results.

• BSP-Net: paper | code | project.
• RfDNet: paper | code | project.