Depth Pro

Depth Pro is a foundation model for zero-shot metric monocular depth estimation that produces high-resolution depth maps with absolute scale, without requiring camera intrinsics or other metadata. It generates a 2.25-megapixel depth map in 0.3 seconds on a V100 GPU. The model uses an efficient multi-scale vision transformer for dense prediction and is trained on a combination of real and synthetic datasets to achieve both metric accuracy and fine boundary detail. It also performs state-of-the-art focal length estimation from a single image.

Key capabilities

• Zero-shot metric depth with absolute scale, no camera intrinsics needed
• High-resolution depth maps with sharp boundaries and high-frequency detail
• Fast inference: 0.3 seconds for a 2.25-megapixel image on a V100 GPU
• Integrated focal length estimation from a single image

Benchmark performance

The model was published at ICLR 2025. In qualitative comparisons, Depth Pro outperforms prior work including Marigold, Depth Anything v2, and Metric3D v2 on boundary sharpness and detail preservation. The reference implementation in this repository has been re-trained and its performance is close to, but does not exactly match, the model reported in the paper.

Evaluation metrics

The model introduces dedicated boundary metrics for depth accuracy. These include scale-invariant boundary F1 for depth-based datasets and scale-invariant boundary recall for mask-based datasets (image matting or segmentation).

Additional details

• Code and weights are released under a custom Apple license.
• Example images for boundary and sharpness comparisons use the AM-2k and DIS-5k datasets.
• For full technical details, see the paper: Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (arXiv 2410.02073, v2 revised April 2025).