# Depth Pro: Sharp Monocular Metric Depth in Less Than a Second

**Authors:** Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, Vladlen Koltun  
**arXiv:** [2410.02073](https://arxiv.org/abs/2410.02073)  
**Date:** 2024-10-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.02073v1/

## Key points

- The paper introduces Depth Pro, a state-of-the-art model for zero-shot metric monocular depth estimation.
- Depth Pro can synthesize high-resolution depth maps with sharpness and intricate details without needing additional metadata like camera intrinsics.
- The model is capable of producing a 2.25-megapixel depth map in just 0.3 seconds on a standard GPU, showcasing impressive speed capabilities.
- Technical contributions include an efficient multi-scale vision transformer, a unique training protocol using real-world and synthetic datasets, and dedicated evaluation metrics for boundary accuracy.
- The researchers also present cutting-edge focal length estimation techniques based on single-image inputs.
- Extensive experimentation shows that Depth Pro outperforms previous works across multiple dimensions.
- The research team has made the code and weights available on GitHub to encourage further exploration and adoption within the scientific community.

## Summary

The paper "Depth Pro: Sharp Monocular Metric Depth in Less Than a Second" by Aleksei Bochkovskii et al. introduces a state-of-the-art model for zero-shot metric monocular depth estimation. The proposed model, Depth Pro, stands out for its ability to synthesize high-resolution depth maps with exceptional sharpness and intricate high-frequency details. Notably, the predictions generated by Depth Pro are metric-based and come with an absolute scale, eliminating the need for additional metadata like camera intrinsics. Moreover, the model showcases impressive speed capabilities by producing a 2.25-megapixel depth map in just 0.3 seconds on a standard GPU. The outstanding performance of Depth Pro can be attributed to several key technical contributions highlighted in the study. These include an efficient multi-scale vision transformer designed for dense prediction tasks, a unique training protocol that leverages both real-world and synthetic datasets to achieve high levels of metric accuracy while ensuring fine boundary tracing accuracy. Additionally, the researchers introduce dedicated evaluation metrics specifically tailored to assess boundary accuracy in estimated depth maps and present cutting-edge focal length estimation techniques based on single-image inputs. Extensive experimentation and analysis conducted by the authors demonstrate that Depth Pro surpasses previous works across multiple dimensions. By releasing both code and weights on GitHub (https://github.com/apple/ml-depth-pro), the research team aims to facilitate further exploration and adoption of their innovative depth estimation model within the scientific community.