Primtive3D: 3D Object Dataset Synthesis from Randomly Assembled Primitives

This REPO is for CVPR2022 Paper: Primitive3D: 3D Object Dataset Synthesis from Randomly Assembled Primitives

We provide two data formats of Primitive3D dataset, namely, PLY data and H5DF data. PLY data contains a set of .ply files, each of which stands for a mesh of a Primitive3D object. Two extra information, source and label, is included in the .ply file, which are the segmentation labels. H5DF data stands for the .h5 files that samples 8192 points for each object. Each .h5 file contains the following keys and data:

• data: coordinates of points with shape N X 8192 X 3.
• normal: normals of points with shape N X 8192 X 3.
• sem_label: point-wise semantic label with shape N X 8192.
• ins_label: point-wise instance label with shape N X 8192.
• p_num: primitive number of the object with shape N.

If ones would like to generate their own datase, please refer to Data Generation. Or one can download the pre-generated dataset with the links in Download Link.

Installation

All experiments have been tested on Python 3.6 and CUDA10.2 environment.

Requirements for Data Generation

Install PyMesh： PyMesh can be install with the '.whl' files in the latest released version 3.0, or our pre-compiled '.whl' for various python versions.

Other dependencies： Other depent packages can be installed by the following command

pip install -r datagen_requirements.txt

Data Generation

To generate Primitive3D dataset in .h5 format, please run:

python gen_primitive3d.py

The dataset generation contains two step:

• generate mesh-based objects with .ply formats in ./data/primitive3d_ply.
• generate point cloud with .h5 format in ./data/primitive3d.h5.

User can change the default path by modify gen_primitive3d.py. Other modification can be done to determine the statistics of dataset generation in this file. The default generated dataset is a small-scale dataset.

Download Link

Name	PLY Link	H5DF Link	Description
Original	link(5.8 GB)	link(36.6GB)	150,000 objects with 1-6 primitives.

If you find our work useful, please considering citing

@inproceedings{li2022primitive3d,
  title={Primtive3D: 3D Object Dataset Synthesis from Randomly Assembled Primitives},
  author={Li, Xinke and Ding, Henghui and Tong, Zekun and Wu, Yuwei and Chee, Yeow Meng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2022}
}