PotSAC: A Robust Axis Estimator for Axially Symmetric Pot Fragments

Abstract

The task of virtually reassembling an axially symmetric pot from its fragments can be greatly simplified by utilizing the constraints induced by the pot's axis of symmetry. This requires accurate estimation of the axis for each sherd, whose 3D data typically contain gross outliers arising from surface artifacts, noisy surface normals and unfiltered data along the break surface. In this work, we propose a simple two-stage robust axis estimator, PotSAC, which is based on a variant of the random sample consensus (RANSAC) algorithm followed by robust nonlinear least squares refinement. Unlike previous work which have either compensated the axis estimation accuracy for robustness against outliers or vice versa, our method can handle the aforementioned outlier sources without compromising its accuracy. This is achieved by carefully designing the method to combine and extend the advantage of each key prior work. Experimental results on real scanned fragments demonstrate the effectiveness of our method, paving the way towards high quality reassembly of symmetric potteries.