Monocular metric depth: How one photo can tell a model distances in meters
A single photo has no built-in ruler. Monocular metric depth models guess distances in meters, and the camera's focal length decides whether those guesses are right. Here is how it works.

Monocular metric depth estimation is the task of looking at a single ordinary photo and predicting how far away every pixel is, in real units such as meters. “Monocular” means one camera, one image. “Metric” means the answer has an absolute scale, so a wall is 3.2 meters away rather than just “farther than the sofa”.
The task is hard because a flat picture throws away distance. A model has to rebuild it from clues, and it also has to know how the camera saw the scene.
Why a photo of a toy car can look like a real one
Picture two photos of a street. One shows a real car 10 meters away. The other shows a toy car 20 centimeters from a lens. With the right framing, both can look almost identical. Nothing in the pixels says which is which.
People solve this with experience: we know roughly how big cars, doors and trees are. Depth models learn the same habit from millions of examples. Early models stopped there and produced only relative depth, an ordering of near and far without units.
Relative depth is enough for a blurred background. It is not enough for a robot that must stop 50 centimeters from a shelf, or for adding a 3D object to a video at the right size.
The missing ingredient is the camera’s lens

Even a model that judges scale well hits a second problem: the lens. A wide-angle lens squeezes a large view into the frame. A telephoto lens shows a narrow slice, so things look bigger and closer.
Engineers describe this with the focal length, here measured in pixels. It links a pixel position to a viewing direction. If the model gets it wrong, every distance is wrong by the same proportion.
Apple’s Depth Pro paper makes this explicit. The network predicts a “canonical inverse depth” map, and metric depth comes from dividing the focal length in pixels by the image width times that map. Depth is proportional to focal length. The Depth Pro 2 README puts it plainly: a focal length 10% too long places everything 10% too far.
Photos often carry the lens value in their EXIF metadata, the data block a camera writes into the file. Scraped web images and many videos do not, or the value is wrong.
How a model learns distance and lens together
Modern systems use a vision transformer, a network that cuts the image into patches and compares them all. Depth Pro, published by Apple researchers Aleksei Bochkovskii, Amaël Delaunoy, Vladlen Koltun and colleagues in October 2024 and presented at ICLR 2025, runs such encoders on patches at several scales. A decoder then fuses them into a depth map. The input is a fixed 1536 by 1536 grid, and the paper reports a 2.25-megapixel depth map in 0.3 seconds on a standard GPU.
Depth Pro adds a separate small convolutional head that predicts the horizontal field of view, the angle the lens covers. The paper says it takes frozen features from the depth network plus features from its own image encoder. It is trained after the depth network, on a data mix that can include images with lens labels but no depth labels. That matters because lens labels are far easier to collect than depth labels.
The result is a model that needs no camera calibration. You can still pass your own focal length, and the model then skips its guess.
What Depth Pro 2 changes and where our news story fits
Depth Pro 2, described in an arXiv paper submitted on 19 November 2026 by Peiyun Hu and seven co-authors, targets video. The abstract says earlier methods trade temporal consistency (depth that does not flicker between frames) against sharp object edges, and claims Depth Pro 2 delivers both together. The paper body was not read for this guide, so its numbers are not repeated here.
The project README lists four depth models: S at 24M parameters, B at 97M, L at 334M and H+ at 872M. All share one 94 MB field-of-view estimator, published on Hugging Face as apple/DepthPro2-focal with 23.5M parameters. For a video, the README says the whole clip gets one focal length: the median of the estimates over its first window of frames.
Our report on the DepthPro2-focal release covers the licensing and packaging details. The weights are limited to non-commercial research.
Where the approach still falls short
The estimate of the lens is itself a guess, and it can be wrong. In the original paper, Depth Pro scored 66.9 on the DDDP set at the 25% tolerance level, against 14.6 for SPEC and 6.8 for UniDepth, with higher being better. Strong numbers on that test do not guarantee accuracy on every phone, crop or edited clip. A cropped or digitally zoomed photo changes the effective focal length, which is one reason EXIF values can mislead.
The authors name a physical limit too. Depth Pro is “limited in dealing with translucent surfaces and volumetric scattering”, such as glass or fog, where a single depth value per pixel is not well defined. The Depth Pro 2 README adds that sky pixels can receive infinite or very large depths.
The public Depth Pro 2 code covers inference and model loading only. Training and evaluation code are not included. The first Depth Pro repository notes that its released model was re-trained, so its performance is close to the paper’s but not identical.
Worth watching: independent tests on real footage from unknown cameras, how the estimator performs against the paper’s claims once readers can check them, and whether the Core ML builds Apple published bring this to apps on Apple silicon.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Daniel ReedAI Research, Safety & Ethics AnalystDaniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…
You may also like

Neural audio codecs: How speech becomes tokens a language model can write

AI scaling laws: Why more compute has meant better models, and where the doubts start

Reward hacking: How AI models learn to win the test without doing the task
