Prediction Error as a Quality Metric for Motion and Stereo

Seventh International Conference on Computer Vision (ICCV'99) |

Published by IEEE Computer Society

This paper presents a new methodology for evaluating the quality of motion estimation and stereo correspondence algorithms. Motivated by applications such as novel view generation and motion-compensated compression, we suggest that the ability to predict new views or frames is a natural metric for evaluating such algorithms. Our new metric has several advantages over comparing algorithm outputs to true motions or depths. First of all, it does not require the knowledge of ground truth data, which may be difficult or laborious to obtain. Second, it more closely matches the ultimate requirements of the application, which are typically tolerant of errors in uniform color regions, but very sensitive to isolated pixel errors or disocclusion errors. In the paper, we develop a number of error metrics based on this paradigm, including forward and inverse prediction errors, residual motion error, and local motion-compensated prediction error. We show results on a number of widely used motion and stereo sequences, many of which do not have associated ground truth data.