Eurecon

Authors: Alexander Morozov, Davide Zgyatti, and Petr Popov, iMolecule, Skoltech.

Equidistant and Uniform Data Augmentation for 3D Objects.

Python library for 3D data augmentation, based on Eurecon algorithm ([1,2]).

• The library works with a variety of 3D file formats, including point cloud formats (.XYZ, .PTS, .PCD) and polygon mesh formats (.STL, .OFF, .OBJ, .PLY, .GLTF).
• Eurecon is computationally efficient, taking ~0.1 seconds to generate 1,000 samples of an object of 1,000 3D points.
• Based on numpy and open3d.
• Simple, flexible API that allows the library to be used in any machine learning pipeline.
• Large, diverse set of transformations based on the amount of input tessellation axes.
• Easy to extend the library to wrap around other libraries.
• Easy to extend to other tasks.

Requirements:

• Python 3.6+
• click==7.0+
• numpy==1.17.2+
• open3d==0.10.0.0+
• tqdm==2.0.0+

Tests were performed on:

• Ubuntu 18.04
• Python 3.8
• click==7.1.2
• numpy==1.17.2
• open3d==0.10.0.0
• tqdm==4.46.1

How to use:

Easiest way to run the augmentation via Eurecon is running eurecon.py with the following command line:

python eurecon.py -r -p -a <path/to/tessellation/axes/file> -i <path/to/input/files> -o <path/to/the/output/directory>

Detailed instructions on how to run Eurecon to reproduce an augmented ModelNet40 dataset are given in augment_ModelNet40.py script.

Example:

python eurecon.py -r 0.5 -p 1 -a resources/tesselation_vertices_layer_0.txt -i examples/test.off -o examples/test_output

How to cite Eurecon:

Morozov A., Davide Z., Popov P., "Equidistant and Uniform Data Augmentation for 3D Objects", IEEE Access, 2022.

References:

[1] Popov, Petr, and Sergei Grudinin. "Rapid determination of RMSDs corresponding to macromolecular rigid body motions." (2014): 950-956.

[2] Popov, Petr, and Sergei Grudinin. "Eurecon: equidistant uniform rigid-body ensemble constructor." Journal of Molecular Graphics and Modelling 80 (2018): 313-319.