Reproduce “PrimeDepth: Efficient Monocular Depth Estimation with a Stable Diffusion Preimage” 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
Denis Zavadski, Damjan Kalšan, Carsten Rother
An efficient monocular depth estimation method extracts a comprehensive "preimage" from a frozen Stable Diffusion model in a single forward pass, achieving up to 100 times faster inference than prior diffusion-based approaches. This method produces more detailed depth maps and shows superior robustness in challenging conditions while maintaining high accuracy compared to other state-of-the-art techniques.
Key insights
- A comprehensive "preimage" extracted from a single forward pass of a frozen Stable Diffusion model is sufficient for highly accurate and efficient zero-shot monocular depth estimation.
- Training with direct pixel-domain loss functions and incorporating an auxiliary semantic segmentation task significantly improves depth estimation performance and detail compared to latent-domain approaches.
- Designing the refiner network with an architectural inductive bias that mirrors the foundation model's decoder effectively integrates multi-scale features, contributing to superior fine-detail capture.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.