QuadricsNet: Learning Concise Representation for Geometric Primitives in Point Clouds

Authors: Ji Wu, Huai Yu, Wen Yang, Gui-Song Xia

This repository contains the official PyTorch implementation of the QuadricsNet paper.

Paper | Video

Introduction

QuadricsNet is an end-to-end framework to learn a concise representation with only 10 parameters for diverse geometric primitives in 3D point clouds. Experiments of primitive parsing on the collected dataset and structure mapping on real-world scenes demonstrate that the quadrics representation is effective and the QuadricsNet framework is robust.

Network Architecture

Qualitative results

Installation

• Basic environment:

conda env create -f environment.yml -n QuadricsNet
conda activate QuadricsNet

• To generate visual results, lightweight MATLAB environment needs to be built. Download MATLAB runtime: Google Drive or Baidu Drive, and then install it:

# Install MATLAB runtime
cd <matlab_runtime_Path>
unzip {matlab_runtime}.zip
./install -mode silent -agreeToLicense yes

# Load {matlab_runtime_install_Path} into an environment variable
export LD_LIBRARY_PATH={matlab_runtime_install_Path} # {matlab_runtime_install_Path} will be printed in the terminal after installation of MATLAB runtime.

# Install quadrics2points function
cd <QuadricsNet_Path>/src/quadrics2points/for_redistribution_files_only/
python setup.py install

Quickstart with the pretrained model

• It is recommended to use at least 128 GB of RAM for more accurate reconstruction.
• Qualitative results are in logs\results_vis.

python test_e2e_vis.py configs/configs_test/config_demo.yml 40000 # 40000 is the number of downsampling points during reconstruction, which can be adjusted based on memory size.

Dataset

The Quadrics dataset is available: Google Drive or Baidu Drive.

• The file structure is as follows:

|-- data
    |-- quadrics_objects
        |-- train
            |-- h5_dense
                |-- 0.h5
                |-- ...
            |-- train.h5
        |-- test
            |-- h5_dense
                |-- 0.h5
                |-- ...
            |-- test.h5
    |-- quadrics_segments
        |-- train
            |-- plane.h5
            |-- ...
        |-- test
            |-- plane.h5
            |-- ...

Training

• Quadrics fitting module training. It is recommended to use at least 48 GB of GPU memory.

bash scripts/train.sh

• Quadrics dection module training. It is recommended to use at least 48 GB of GPU memory.

python train_detection.py configs/configs_train/config_detection.yml

• [Optional] End-to-end training. It is recommended to use at least 96 GB of GPU memory.

python train_e2e.py configs/configs_train/config_e2e.yml

Test

python test_e2e_vis.py configs/configs_test/config_e2e.yml 40000 # To use the retrained model, fitting_model_path and detection_model_path in config_e2e.yml need to be modified.

Citation

If you find our work useful, please consider citing our paper:

@article{quadricsnet,
  title={QuadricsNet: Learning Concise Representation for Geometric Primitives in Point Clouds},
  author={Wu, Ji and Yu, Huai and Yang, Wen and Xia, Gui-Song},
  journal={arXiv preprint arXiv:2309.14211},
  year={2023}
}