Landmark
A numbered reference point at a fixed spot
- A landmark is a reference point where the target is fixed and every point carries a number set in advance. Point number twelve means the same spot in any photo.
- Because the numbering is fixed, matching numbers can be compared directly. That's what makes landmarks useful.
- Their first job is lining things up. Tilted, differently sized inputs get rotated and stretched to match the reference points before the next step gets easier.
- How many points get used depends on the purpose. Some lists run a sparse few dozen, and others lay down a dense net of several hundred.
- What matters isn't how a point looks but the distance and ratio between points — that's the raw material for judging whether an eye is closed or a mouth is open.
Contents
1The analogy
Mountaintops and roadsides across a country are dotted with official survey markers. They look like plain metal discs, but each one carries its own ID number, and its coordinates are written down in a registry. Which marker sits where doesn't change even after ten years.
Whenever land gets resurveyed or a map redrawn, the markers get found first. Measuring your own patch perfectly means nothing without a shared reference — other people's measurements won't line up with yours. Anchor a few markers, though, and readings taken by different people on different days snap onto a single map.
A landmark is that survey marker. Whatever photo comes in, point number so-and-so always points to that same fixed spot, and everything else gets aligned against those anchors.
2In detail
The number is the location
What matters about a landmark isn't how many points there are — it's that a number and a spot are paired and locked together. Define the list once, and in any input, the same number points to the same part. Point six might always mean the outer corner of the left eye; point fourteen, the middle of the lower lip.
That agreement is what makes photos directly comparable. Point twelve in today's photo and point twelve in last week's photo mark the same part, so subtracting one value from the other shows how far that part moved. Without numbering, you'd have to solve "which point matches which point" all over again first.
Building the list is a human job. A training set of thousands of photos gets hand-marked with numbered points first, and the model learns from that where to place the same numbered point on a new photo. Two tools built on different lists have numbers that don't line up, so results can't be swapped between them.
They line things up first
A landmark's first task isn't reading meaning — it's putting the input in order. Incoming photos arrive tilted, near or far, and off to one side of the frame. Nothing can be compared like that.
So a handful of reference points get anchored. The two points standing for the eyes, say, get pinned to a fixed spot on the frame, and the photo gets rotated and resized to match. Once that's done, different photos sit inside the same frame, and everything computed afterward gets far more stable.
This is part of why face recognition holds up so well against changes in angle and distance — not because the recognition step itself is so skilled, but because landmarks lined things up before it even started. Video-call touch-up filters and virtual glasses try-on both run on top of this same alignment.
How many points depends on the purpose
A sparse list runs a few dozen points — enough to mark the borders of the eyes, nose, mouth, and jawline. It's light to compute and plenty for lining up a face or roughly sorting expressions. Older devices and tools running in a browser tend to use this level.
A dense list can run to several hundred. Connect the points into triangles and they form a mesh over the surface; lay a pattern on top and it's what makes glasses or a mask look like they're stuck to a moving face. More points track the surface's curves more finely, but the computation gets heavier and each point jitters more too.
More than the purpose calls for is wasted effort. Using a several-hundred-point mesh just to tell whether an eye is closed burns extra battery without improving the answer.
Meaning comes from the gaps between points
A landmark by itself is just a coordinate on a list. The way meaning gets pulled out is by measuring the distance between two specific numbered points. A shrinking gap between the points standing for the upper and lower eyelid reads as an eye closing; a widening gap between upper and lower lip reads as a mouth opening.
Using raw distance is risky, though — moving closer to the camera makes every distance grow at once. So distances get expressed as a ratio against another distance instead. Use the gap between the two eyes as the ruler, and the result stops swinging with how near or far the camera is.
A handful of these ratios can support a lot of judgments. Drowsy-driving alerts watch for the eyelid ratio staying narrow past a set time; a video call's "who's speaking" indicator watches the lip ratio rise and fall. It's closer to the truth to say a few length ratios are being tracked than that expressions are truly being read.
3More precisely
A single landmark is expressed as a horizontal and vertical position plus a confidence value, and some tools add depth on top. Position values often come out as a fraction of the frame size, so they still hold up when the photo's dimensions change. Rather than pointing directly at a spot, the usual computation lays down a rough layout first, then nudges each point a little at a time to refine it.
Landmark is a branch of keypoint. Both mark a location with a point — where they split is that a landmark has a fixed target and a tightly fixed numbering. A short, purpose-shifting list, like one for joints, usually gets called a keypoint; a list that's hardened into a standard, like one for a face, gets called a landmark. The line isn't crisp, so tool documentation sometimes mixes the two words.
The analogy breaks down somewhere too. A survey marker is physically driven into the ground, so finding it settles the matter. A landmark isn't marked anywhere at all — it's a position guessed fresh each time, imitating where a human labeled it in training data, and it lands somewhere wrong when the spot is covered or the angle is severe. A marker's coordinates are a settled value in a registry; a landmark's coordinates always carry some error.
4Try it yourself
5Common misconceptions
It's easy to think landmark and keypoint are two different technologies, but actually landmark refers to the branch of keypoints where the numbering and the target are fixed.
It's easy to think more points means more accurate, but actually as many as the purpose needs is enough, and more points bring more computation and more jitter along with them.
It's easy to think landmarks alone reveal expression or emotion, but actually they only show that a few distance ratios changed — they don't read what's actually going on inside.
7One-line summary
In shortA landmark is a reference point numbered in advance, and because that numbering stays fixed, different photos can be measured against the same ruler.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02