Metric Depth Estimation from Arbitrarily Degraded Low-Resolution Depth Prompts
AdaDS is a generalizable framework for depth super-resolution that recovers high-resolution depth maps from low-resolution inputs. It uses Gaussian smoothing contraction properties in diffusion processes to adaptively select a starting refinement timestep based on uncertainty, inject tailored noise, and leverage a pre-trained diffusion model for robust recovery across various degradation levels.