Representing Point Clouds with Generative Conditional Invertible Flow Networks

This is a PyTorch implementation of the paper:

Representing Point Clouds with Generative Conditional Invertible Flow Networks
Michał Stypułkowski, Kacper Kania, Maciej Zamorski, Maciej Zięba, Tomasz Trzciński, Jan Chorowski
Preprint. Under review.

Introduction

This paper focuses on a novel generative approach for 3D point clouds that makes use of invertible flow-based models. The main idea of the method is to treat a point cloud as a probability density in 3D space that is modeled using a cloud-specific neural network. To capture the similarity between point clouds we rely on parameter sharing among networks, with each cloud having only a small embedding vector that defines it. We use invertible flows networks to generate the individual point clouds, and to regularize the embedding vectors. We evaluate the generative capabilities of the model both in qualitative and quantitative manner.

Requirements

Stored in requirements.txt.

Training

Run the training process with:

python experiments/train/train_model.py --config configs/cif_train.yaml

You can also download pretrained models for:

• airplanes,
• cars,
• chairs.

Evaluation

Run python experiments/test/EXPERIMENT_NAME.py --config configs/cif_eval.yaml, where EXPERIMENT_NAME can be on of the following:

1. train_reconstruction to reconstruct the training set
2. test_reconstruction to reconstruct the test set
3. sampling to sample new objects
4. interpolation to interpolate between shapes in latent space
5. common_rare to find most casual and unique point clouds in the dataset
6. metrics_eval to evaluate performance of the model. You need to install pytorch_structural_losses from here. Please note that it calculates only Coverage and MMD. For full evaluation we used PointFlow's script.

Rendering

1. Install docker container of Mitsuba Renderer as:

$ docker build -t mitsuba <path-to-downloaded-repository-of-mitsuba>
$ docker run -it -p 8000:8000 --name mitsuba mitsuba:latest 

This will start a service for rendering purposes.

2. Run:

$ python utils/visualize_points.py \
    <input-file> \
    <output-folder-or-file> \
    [--torch] \
    [--rotated] \
    [--batch] \
    [--port <port>]

where:

• <input-file> is a path to either *.npy file or a file that can be depickled by torch. The file contains points as N x 3 matrix or B x N x 3 where B is a batch size and N - number of points. If the file has 3 dimensions, you need to use --batch as well.
• <output-folder-or-file> is a directory where renders from a batch should saved as 1.png, 2.png, ..., <B>.png. If --batch is not used then it should be a file path, for example: ~/output/img.png.
• --torch optional flag pointing that the <input-file> should be depickled with torch.load
• --rotated optional flag that rotates point clouds prior rendering. It should be used in cases where the rendered shape is rotated.
• --port <port> is a port for the mitsuba service if it was run with other port than 8000.

Citation

@article{stypulkowski2020cif,
  title={Representing Point Clouds with Generative Conditional Invertible Flow Networks},
  author={Stypu{\l}kowski, Micha{\l} and Kania, Kacper and Zamorski, Maciej and Zi{\k{e}}ba, Maciej and Trzci{\'n}ski, Tomasz and Chorowski, Jan},
  journal={arXiv},
  year={2020}
}