Fig. 9.

Depth error and bad3 metric (percentage of disparities $>$3 pixels error) for multiple manual alignments. Results with a bad3 of $>$20% for both HSM (level 1) and DeepPruner are considered as outliers due to human error. Outliers were not taken into consideration when computing the average manual alignment (in pink). This average produces a consistently low bad3 error for the best performing network and the depth error is around 2 mm for all but the most challenging images.