Code of paper "Noisy Dual Principal Component Pursuit", ICML 2019

Synthetic Experiments

• Requirements

  • Matlab
  • C++
  • Armadillo (C++ scientific computing library)
  • Python3
  • numpy, matplotlib
  • [optional] OpenBLAS

  Basically, we use MATLAB for simple tasks, C++ for heavy jobs and Python to do some plotting after obtaining data through MATLAB/C++ programs.

• Usage (tested under Mac OS)

  • Matlab program can be executed directly
  • Generally, each .cpp file is paired with a driver.py file (just run the driver file is enough)
  • An installation of OpenBLAS will further accelerate the C++ programs but the compilation in driver.py needs to be changed accordingly

3D Roadplane Estimation

• demo.m is a toy example that runs single subspace learning algorithms on real 3D road plane detection data. Once we have selected the frame and click run button, it instantly runs the algorithms and returns the clustering metrics, geometric metrics and algorithmic metrics as mentioned in the paper. Also, a poster showing the projections of the separated point clouds onto the image is generated after the execution of the program.

• /data is a folder containing annotations for point clouds and corresponding images.

• /algorithms is a folder containing various single subspace learning algorithms.

Citation

If you find the code or results useful, please cite the following paper:

@inproceedings{ding2019noisy,
    title={Noisy dual principal component pursuit},
    author={Ding, Tianyu and Zhu, Zhihui and Ding, Tianjiao and Yang, Yunchen and Vidal, Rene and Tsakiris, Manolis  and Robinson, Daniel},
    booktitle={Proceedings of the International Conference on Machine learning},
    pages={1617--1625},
    year={2019}
}