0

Does it make sense to use RANSAC-type algorithms (RANSAC, MSAC, MLESAC, etc.) for small data sets (20-30 points)?

On the one hand, all the points need to be accounted for and this can be done with other robust methods (robust regression, expectation maximization). On the other hand, RANSAC-like algorithms may offer greater computational stability, which is especially important when the calculation need to be performed repeatedly.

Tim
  • 108,699
  • 20
  • 212
  • 390
Roger Vadim
  • 1,481
  • 6
  • 17

1 Answers1

1

The Why isn't RANSAC most widely used in statistics? thread seems to respond to your concerns.

If the dataset is small, then the problem with RANSAC is that it will throw away data, so your model will be fitted to even smaller data. If you don't have much data, throwing away any of it, may not be the best approach.

Tim
  • 108,699
  • 20
  • 212
  • 390
  • After a hundred of iterations I will have high chance of all the data points covered. On the other hand, using RANSAC for robust regression may be numerically more stable than, e.g., iteratively re-weighted least squares. – Roger Vadim Dec 16 '19 at 17:54
  • @Vadim "all data points covered" to choose the subset of points to be used to train the final model, so to train it you will throw away data, that's the point of RANSAC. – Tim Dec 16 '19 at 18:49