Abstract
We formulate single-view 3D super-resolution as a structured inverse problem over an image-space coordinate field, designed for unguided scenarios where high-resolution RGB guidance is unavailable (e.g., low-light or privacy-sensitive environments). Our framework, GeometRep-SR, maps low-resolution 3D observations into a bijective 2D representation in which standard image super- resolution networks can perform geometrically consistent inference. We instantiate this formulation using the Projected Normalized Coordinate Code (PNCC), which represents the visible 3D surface as a per-pixel XYZ coordinate field aligned with camera rays. By lifting scalar depth into this 3-channel coordinate space, we allow standard 2D backbones to learn explicit spatial coherency, effectively substituting RGB guidance with geometric self-guidance. We evaluate GeometRep- SR with two implementations: using Swin Transformers for high accuracy, and using Vision Mamba (VM) for high efficiency. Crucially, our Swin Transformer implementation surpasses state-of-the-art RGB-guided methods (e.g., SGNet) in accuracy, proving that explicit geometric encoding can replace high-resolution RGB priors. Simultaneously, the Vision Mamba variant delivers competitive performance while being 18× faster, enabling real-time unguided 3D perception. This establishes our pipeline as a robust solution for real-world 3D perception.
