Efficient Learning on Point Clouds with Basis Point Sets

Update: pure PyTorch implementation of the BPS encoding is now available, thanks to Omid Taheri.

Basis Point Set (BPS) is a simple and efficient method for encoding 3D point clouds into fixed-length representations.

It is based on a simple idea: select k fixed points in space and compute vectors from these basis points to the nearest points in a point cloud; use these vectors (or simply their norms) as features:

The basis points are kept fixed for all the point clouds in the dataset, providing a fixed representation of every point cloud as a vector. This representation can then be used as input to arbitrary machine learning methods, in particular it can be used as input to off-the-shelf neural networks.

Below is the example of a simple model using BPS features as input for the task of mesh registration over a noisy scan:

FAQ: what are the key differences between standard occupancy voxels, TSDF and the proposed BPS representation?

continuous global vectors instead of simple binary flags or local distances in the cells;
smaller number of cells required to represent shape accurately;
BPS cell arrangement could be different from a standard rectangular grid, allowing different types of convolutions;
significant improvement in performance: simply substituting occupancy voxels with BPS directional vectors results in a +9% accuracy improvement of a VoxNet-like 3D-convolutional network on a ModelNet40 classification challenge.

Check our ICCV 2019 paper for more details.

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{prokudin2019efficient,
  title={Efficient Learning on Point Clouds With Basis Point Sets},
  author={Prokudin, Sergey and Lassner, Christoph and Romero, Javier},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  pages={4332--4341},
  year={2019}
}

Usage

Requirements

Python >= 3.7;
scikit-learn >= 0.21;
tqdm >= 4.3;
PyTorch >= 1.3 (for running provided demos)
trimesh, pillow, pyntcloud (for running human mesh registration code)
psbody-mesh (to compute FAUST registrations correspondences)

Installation

pip3 install git+https://github.com/sergeyprokudin/bps

Code snippet

Converting point clouds to BPS representation takes few lines of code:

import numpy as np
from bps import bps

# batch of 100 point clouds to convert
x = np.random.normal(size=[100, 2048, 3])

# optional point cloud normalization to fit a unit sphere
x_norm = bps.normalize(x)

# option 1: encode with 1024 random basis and distances as features
x_bps_random = bps.encode(x_norm, bps_arrangement='random', n_bps_points=1024, bps_cell_type='dists')

# option 2: encode with 32^3 grid basis and full vectors to nearest points as features
x_bps_grid = bps.encode(x_norm, bps_arrangement='grid', n_bps_points=32**3, bps_cell_type='deltas')
# the following tensor can be provided as input to any Conv3D network:
x_bps_grid = x_bps_grid.reshape([-1, 32, 32, 32, 3])

Demos

Clone the repository and install the dependencies:

git clone https://github.com/sergeyprokudin/bps
cd bps
python setup.py install
pip3 install torch h5py

Check one of the provided examples:

ModelNet40 3D shape classification with BPS-MLP (~89% accuracy, ~30 minutes of training on a non-GPU MacBook Pro, ~3 minutes of training on Nvidia V100 16gb):

python bps_demos/train_modelnet_mlp.py

ModelNet40 3D shape classification with BPS-Conv3D (~92% accuracy, ~120 minutes of training on Nvidia V100 16gb):

python bps_demos/train_modelnet_conv3d.py

FAUST body mesh registration: check tutorial notebook to run the evaluation of the pre-trained model.

You can directly download the results (predicted alignments, computed correspondences, demo video) here.

Results are also visualised in this video.

Running pre-trained model on your own 3D body scans: download the model checkpoint (mirror):

mkdir data
cd data
wget --output-document=mesh_regressor.h5 https://www.dropbox.com/s/u3d1uighrtcprh2/mesh_regressor.h5?dl=0

Run the model, providing the paths to your own *.ply file and output directory. You can test that everything works by running the following synthetic example:

cd bps_demos
python run_alignment.py demo_scan.ply ../logs/demo_output

If a directory is provided as a first parameter, the alignment model will be ran on all *.ply files found.

Here is an example of a prediction on some noisy real scan:

License

This library is licensed under the MIT-0 License. See the LICENSE file.

Note: this is the official fork of the Amazon repository written by the same author during the time of internship. Latest features and bug fixes are likely to appear here first.