Depth Pro

Apple’s open-source monocular metric-depth model that outputs a sharp depth map in meters from a single image.

Depth Pro was proposed by Bochkovskii, Koltun, and colleagues at Apple; the paper, code, and weights were released in October 2024, and it was published at ICLR 2025. It does zero-shot monocular metric depth estimation: given only a single image, with no metadata like camera intrinsics required, it outputs absolute depth in meters while also estimating the camera’s focal length. It’s built for sharp edges and fine detail, and the paper reports generating a 2.25-megapixel depth map in 0.3 seconds on an ordinary GPU. Technically, it uses an efficient multi-scale vision Transformer for dense prediction, trains on a mix of real and synthetic data, and introduces a dedicated metric for measuring the accuracy of depth boundaries. Relative depth only tells you what’s nearer or farther; metric depth carries true scale, which is what lets it be back-projected directly into a point cloud a robot can use.

ExampleGiven a kitchen photo downloaded from the web with no capture parameters, Depth Pro outputs how many meters each pixel is from the camera along with an estimated focal length, letting a true-scale point cloud be recovered with a pinhole camera model.

Also called

Apple Depth Pro, Depth Pro: Sharp Monocular Metric Depth in Less Than a Second

Related

Monocular Depth Estimation · Metric Depth / Relative Depth · Depth Anything · Metric3D · MoGe · Camera Intrinsics

Sources

Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (arXiv 2410.02073)
apple/ml-depth-pro (GitHub)

As of

2025-04