Evaluation of Direct Plane Fitting for Depth and Parameter Estimation

Abstract

Recently, a model-based depth estimation technique has been proposed, which estimates surface model parameters by means of Hooke-Jeeves optimization. Assuming a parametric surface model, the parameters best explaining the perspective changes of the surface between different views are estimated. This constitutes a fitting of models directly into stereo images, which is in contrast to the usual approach of fitting models into pre-processed disparity data. In this paper, we conduct a comparison of the image fitting based on Hooke-Jeeves, an image fitting based on gradient descent and a disparity fitting based on RANSAC. We show that the image fitting based on Hooke-Jeeves as well as the image fitting based on gradient descent are sensitive to occlusion. However, we also propose a simple pre-processing that eliminates this problem. Our experiments revealed that all three approaches have a similar depth accuracy. However, tests under challenging conditions show that the fitting based on Hooke-Jeeves is more robust than RANSAC and gradient descent.