Core Concepts
A method for estimating robust disparity confidence intervals in stereo matching problems using possibility distributions.
Abstract
The article presents a method for creating robust disparity confidence intervals in stereo matching problems. The key highlights are:
The method relies on possibility distributions to model the epistemic uncertainty in the cost volume, interpreting it as an expert's opinion on the similarity of image patches.
Confidence intervals are deduced from the α-cuts of the possibility distributions, providing a lower and upper bound on the disparity for each pixel.
The method is designed to be integrated into classical 3D reconstruction pipelines using cost-volume based stereo matching algorithms.
Post-processing steps like sub-pixel refinement and filtering are handled to maintain consistency between the disparity map and the confidence intervals.
A statistical regularization is applied in low-confidence areas to extend the intervals and account for potential biases in the cost curves.
The method is evaluated on the Middlebury stereo datasets and a dataset of satellite images, demonstrating high accuracy (over 90%) and reasonably small interval sizes, especially in high-confidence areas.
The authors argue that providing confidence intervals, in addition to the usual confidence measures, gives a deeper understanding of the uncertainty in stereo matching and can be useful for downstream applications like 3D reconstruction from satellite imagery.
Stats
The disparity range for the Middlebury datasets is between 60 and 1110 pixels.
The disparity range for the satellite images is between 20 and 50 pixels.