The author proposes an architecture that leverages high-resolution ground truths to improve prediction quality while using low-resolution inputs efficiently.
