This convolution has two effects: First, radiometric offsets are removed. Second, the Sx eliminates y gradients
or, respectively, horizontal structures that might have a negative impact in matching in case of (unavoidable but
small) errors in the image orientation. However, our main driver for switching the cost from Mutual Information
(MI) that we initially used for ADS processing (cp. Gehrke et al., 2010) has been occasional errors; in brief: MI cost
reflects the probability of a certain pair of intensities to be a valid match, and for a particular project in North West
Geomatics’ production the probability computation was dominated by grassland and, applied to the neighboring
road (with similar intensities but slightly different pair matches due to a different BRDF), elongated outliers
occurred despite the SGM smoothness constraints. Already used for large projects of various type – including all
examples presented in this paper –, the Sobel-based cost has not yet shown any issues on flat structures and open
areas in general. This is crucial when aiming for the derivation of a DEM.