Depth from one photo is relative, not absolute

Monocular depth estimation is a real, working technique, and it is routinely mistaken for something it is not. A model can look at one photograph and produce a plausible map of what is nearer and what is farther. It cannot tell you how large anything actually is, because that information was never in the photograph to begin with.

What a depth model actually produces

Trained on large datasets of photos paired with true depth (usually gathered by a separate sensor, or inferred from video and multiple viewpoints during training), a monocular depth model learns to associate visual cues - blur, texture gradient, occlusion, familiar object scale - with relative distance from the camera. Given a new single photo, it outputs a depth map: a value per pixel describing roughly how far that point is from the lens, relative to the rest of the scene.

This is genuinely useful. It is also, crucially, relative. The map tells you that one point is nearer than another by some proportion. It does not, and structurally cannot, tell you the actual distance in centimetres, because nothing in a single 2D image fixes the scale of the scene it depicts.

Why scale stays missing

A photograph contains no scale information by construction - the same framing can result from a small subject held close or a larger one held farther away, and the pixels are identical either way. Depth estimation does not fix this; it operates entirely within the same scale-ambiguous world. A depth model can correctly say "this point is twice as far from the camera as that one" while having no way to say whether the units involved are centimetres or metres, because both readings produce the exact same relative map.

Some systems resolve this using external cues - a second camera (stereo, which genuinely does recover scale from the baseline between two lenses), a known reference object in frame, or a depth sensor bolted onto the phone. None of these are "monocular" depth from a single ordinary photo anymore; they are a different measurement setup that happens to also produce a depth-like output. The scale-ambiguity problem is specific to inferring depth from one image alone, which is the situation almost every uploaded photo is actually in.

There is a further wrinkle even within monocular estimation: many depth models are themselves trained to predict relative depth only, because that is what most training data reliably supports, and a smaller subset are trained toward metric depth using datasets where true distances were captured by an accompanying sensor. A metric-depth model still needs the training distribution to resemble the input at inference time; point it at an unfamiliar framing or distance and its absolute-scale output degrades toward the same relative-only guess a purely ordinal model would give.

What this settles and what it does not

Depth estimation genuinely improves a model's read on shape and contour - it can distinguish "this is a bump because it's physically nearer the camera" from "this is a bump because of shading," which flat 2D texture analysis cannot always do on its own. That is a real improvement to how well a model reads form and proportion from a photo.

It does nothing for the question of absolute size, which stays exactly as unanswerable from a single photo as it was without any depth model in the pipeline. A reference object placed in frame does not get picked up automatically either, unless a system was specifically built to detect and use one, which most were not.

If an actual figure in centimetres is what you want, the honest route stays a tape and a repeatable method rather than any amount of depth modelling from a photo - Measure My Cock covers that method directly, and it is worth reading there rather than expecting a scoring tool to quietly solve it. Rate Cock does not present its size axis as a physical measurement for this reason, and a commissioned human reviewer runs into the identical wall from the opposite direction - a person looking at a photo has no more access to true scale than a depth model does, for the same underlying reason. Reading what any of these tools' numbers do and do not claim is worth doing before comparing them to each other, and that habit is covered from the user's side elsewhere.