Landmark detection, and what it adds to a rubric

A keypoint model does a narrower job than a scorer. Instead of producing a single number, it marks specific points on the subject - a set of coordinates locating particular anatomical landmarks - and hands those coordinates downstream, either to a scoring head that turns them into features or directly to a formula that computes ratios between them. This is the same family of technique used in pose estimation, where a model marks joints on a body rather than describing the pose in prose. Face landmark models, used for everything from filters to face-swap detection, are the most familiar public example, and they work on exactly the same principle applied to a different set of points.

How a keypoint model works

Keypoint detection is trained on images labelled with the exact pixel location of each point of interest, and the model learns to predict a heatmap for each point - a probability distribution over where that point is likely to be, peaking at the true location once training has converged. The final coordinate is usually the peak of that heatmap. This is a different learning target from a scoring model's: a scoring model is trained to match a human-assigned number, while a keypoint model is trained to match a human-marked location, which is a more constrained and in some ways easier task, since a location is objectively checkable against the label in a way a subjective score is not.

What ratios buy you

Once a handful of points exist as coordinates, distances and ratios between them can be computed directly - length in pixels between two points, the ratio of one span to another, symmetry of one side against the other. These are geometry-like features in a way a raw embedding vector is not: they can be inspected, plotted, and explained in a sentence, which is a real advantage over a score that emerges from an opaque vector with no interpretable steps in between. A pipeline that reports "these two points measured this many pixels apart, at this ratio to these other two points" is showing its work in a way a black-box score never does, and that legibility is the actual value keypoint-based features add to a rubric.

Why it is still not measurement

Every quantity a keypoint model produces is in pixels, on a two-dimensional projection, with no reference to physical scale anywhere in the calculation. A ratio between two pixel distances is scale-invariant in a useful way - it does not care how far the camera was from the subject, because both distances shrink or grow together - which is genuinely useful for symmetry and proportion. An absolute pixel distance is not scale-invariant at all, and converting it to a physical unit would require a scale reference in the frame, which training never gives a model any way to use unless the pipeline was specifically built around a reference object. Keypoint ratios are a more legible feature than an unexplained embedding, and they remain exactly as blind to actual size as everything else in the pipeline described for what a model sees. If a physical figure in centimetres is what is actually wanted, Measure My Cock's tape-and-method approach is the honest route to it, not a more sophisticated keypoint model.

Where landmark models fit in a pipeline

Keypoint detection usually runs after detection has cropped the subject and before the final scoring head, feeding derived ratios into the model alongside, or instead of, the raw embedding. It is a design choice rather than a universal step - many pipelines skip it entirely and score straight from the embedding, and the ones that add it are usually trying to make the shape and proportion axes more explainable rather than more accurate. Rate Cock uses landmark-derived ratios as one input to its shape axis specifically for that reason. A human judge does something loosely analogous by eye without ever formalising points and coordinates, which is one of several differences between what a person notices and what a model can be made to show its work on. Reading a shape or proportion score with this mechanism in mind is worth doing before treating it as more precise than a holistic one; Penis Rater's notes on interpreting per-axis scores cover that from the user's side. The legibility is real even when the underlying number is not, and that combination - a number you can audit, still built on a projection with no depth in it - is closer to what keypoint models actually offer than either "measurement" or "black box" would suggest.