Abstract
Dense stereo matching takes overlapping image pairs as input and outputs a disparity map which encodes
pixel-by-pixel matches between the images. Recently, there has been an interest in ranking the quality, or
even quantifying the accuracy, of disparity estimates. The proposed methods can be described as either
uncertainty estimators or confidence metrics. Uncertainty estimators are a small minority of the research.
However, they have the potential to be the most useful because they estimate disparity accuracy (in pixel
units) that can be used to threshold matches or carried forward using error propagation. The majority of the
research deals with confidence metrics which give an ordinal (or binary) ranking of a match’s quality relative
to other matches. Confidence metrics do not have units and thus are useful primarily for thresholding matches
from mismatches. The methods could also be described as handcrafted or deep-learning based. The majority
of the research focused on outdoor driving scenes. Hence, our interest–application to a satellite semi-global
matching pipeline–is a domain shift that may challenge deep-learning based methods. We conclude by
recommending five handcrafted and two deep-learning based methods for evaluation in our pipeline.
