Reproduce “Zero-Shot Metric Depth with a Field-of-View Conditioned Diffusion Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Saurabh Saxena, Junhwa Hur, Charles Herrmann, Deqing Sun, David J. Fleet
The Diffusion for Metric Depth (DMD) model from Google DeepMind and Google Research introduces a field-of-view conditioned diffusion approach, achieving new state-of-the-art zero-shot monocular metric depth estimation. It significantly improves generalization across diverse indoor and outdoor scenes, demonstrating up to a 33% reduction in relative error on zero-shot outdoor datasets compared to prior methods.
Key insights
- Generic denoising diffusion models, when properly conditioned, can achieve superior zero-shot metric depth estimation compared to specialized architectures.
- Representing depth in log-scale effectively models the diverse depth ranges of both indoor and outdoor environments within a single model.
- Explicitly conditioning on the camera's Field-of-View (FOV) combined with FOV augmentation is essential for resolving depth-scale ambiguity and achieving robust generalization to unseen camera intrinsics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.