This article presents a new approach to learn geometries such as depth and surface normal from images. The introduced method, called Adaptive Surface Normal (ASN) constraint, extracts geometric context that encodes geometric variations present in the input image. It correlates depth estimation with geometric constraints and prioritizes regions that exhibit significant geometric variations. This allows for the generation of high-quality 3D geometry from images. The method is validated against state-of-the-art methods on diverse indoor and outdoor datasets, demonstrating its efficiency and robustness.
Publication date: 9 Feb 2024
Project Page: https://github.com/xxlong0/ASNDepth
Paper: https://arxiv.org/pdf/2402.05869
