Bucketing Techniques in Robust Regression for Computer Vision

Abstract

Robust parameter estimation methods have become very popular in the computer vision community. Nevertheless, both optimization models and resolution algorithms coming from robust statistics must be adapted to correctly tackle the specificities of visual data. Among these adapted techniques, computer-vision researchers frequently use bucket-based partitions of the data (bucketing techniques). This work points out the key ideas and features of bucketing techniques. A new stochastic sampling scheme is proposed and defended. We also try to answer several questions, which are generally -and perhaps voluntarily-bypassed : “does the bucketing strategy influence the regression process?”; “how should the data be split into buckets to get the best fits both numerically and physically?” . . .

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.