Based our performance data for all backbones, we analyze to which degree performance can be attributed to general properties of the backbone (input image resolution, feature dimension, number of parameters). We find strong relationships with all properties for semantic segmentation and depth.
