The Hypersim Dataset For many fundamental scene understanding tasks, it is difficult or impossible to obtain per-pixel ground truth labels from real i