apple-aiml-research/ml-depth-pro

What it solves

Depth Pro addresses the challenge of estimating accurate, high-resolution metric depth from a single image (monocular depth estimation) without needing camera metadata like intrinsics. It specifically solves the trade-off between speed, resolution, and sharpness, providing absolute scale predictions with fine boundary details.

How it works

The system uses a foundation model based on an efficient multi-scale vision transformer designed for dense prediction. It was trained using a protocol that blends real and synthetic datasets to ensure both high metric accuracy and precise boundary tracing. The model can estimate focal length from a single image and produces a 2.25-megapixel depth map in approximately 0.3 seconds on a standard GPU.

Who it’s for

This tool is for researchers and developers working in computer vision, robotics, or 3D reconstruction who need fast, zero-shot metric depth estimation from standard images.

Highlights

• Zero-shot metric depth: Predicts absolute scale without requiring camera intrinsics.
• High speed: Generates 2.25-megapixel depth maps in under 0.3 seconds on a GPU.
• Sharp details: Focuses on high-frequency details and fine boundary tracing.
• Integrated evaluation: Includes dedicated metrics for measuring boundary accuracy in depth maps.