Robust estimators

[mpizenberg]

Black 1996 uses Geman-McClure robust M-estimator.

@article{black1996robust,
  title={The robust estimation of multiple motions: Parametric and piecewise-smooth flow fields},
  author={Black, Michael J and Anandan, Paul},
  journal={Computer vision and image understanding},
  volume={63},
  number={1},
  pages={75--104},
  year={1996},
  publisher={Elsevier}
}

Zhang 1997 presents parameters estimation from noisy data, including robust techniques
such as M-estimators and least median of squares (but not Ransac!).

@article{zhang1997parameter,
  title={Parameter estimation techniques: A tutorial with application to conic fitting},
  author={Zhang, Zhengyou},
  journal={Image and vision Computing},
  volume={15},
  number={1},
  pages={59--76},
  year={1997},
  publisher={Elsevier}
}

Stewart 1999 presents usage of M-estimators and least median squares (LMS)
and RANSAC in the context of computer vision.

@article{stewart1999robust,
  title={Robust parameter estimation in computer vision},
  author={Stewart, Charles V},
  journal={SIAM review},
  volume={41},
  number={3},
  pages={513--537},
  year={1999},
  publisher={SIAM}
}

In part 2 of Baker 2003, the authors show how a robust M-estimator can be used
within an inverse compositional iteratively reweighted least squares (IRLS) algorithm

@inproceedings{baker2003lucas,
  title={Lucas-kanade 20 years on: A unifying framework: Part 2},
  author={Baker, Simon and Gross, Ralph and Ishikawa, Takahiro and Matthews, Iain},
  booktitle={International Journal of Computer Vision},
  year={2003},
  organization={Citeseer}
}

In Audras 2011, the authors use a robust M-estimator

@inproceedings{audras2011real,
  title={Real-time dense appearance-based SLAM for RGB-D sensors},
  author={Audras, Cedric and Comport, A and Meilland, Maxime and Rives, Patrick},
  booktitle={Australasian Conf. on Robotics and Automation},
  volume={2},
  pages={2--2},
  year={2011}
}

Klose 2013 uses robust M-estimator Huber and Tukey.
Removes median abs res, and estimate std before computing weights.
Also introduce global affine illumination changes.

@inproceedings{klose2013efficient,
  title={Efficient compositional approaches for real-time robust direct visual odometry from RGB-D data},
  author={Klose, Sebastian and Heise, Philipp and Knoll, Alois},
  booktitle={2013 IEEE/RSJ International Conference on Intelligent Robots and Systems},
  pages={1100--1106},
  year={2013},
  organization={IEEE}
}

In Kerl 2013, they use a t-distribution instead of M-estimator.

@inproceedings{kerl2013robust,
  title={Robust odometry estimation for RGB-D cameras},
  author={Kerl, Christian and Sturm, J{\"u}rgen and Cremers, Daniel},
  booktitle={2013 IEEE international conference on robotics and automation},
  pages={3748--3754},
  year={2013},
  organization={IEEE}
}

Gutierrez 2015 evaluate huber, tukey and Student t-distribution. Find that t-dist is best and more stable.

@inproceedings{gutierrez2015inverse,
  title={Inverse depth for accurate photometric and geometric error minimisation in RGB-D dense visual odometry},
  author={Guti{\'e}rrez-G{\'o}mez, Daniel and Mayol-Cuevas, Walterio and Guerrero, Josechu J},
  booktitle={2015 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={83--89},
  year={2015},
  organization={IEEE}
}