Key points
- A keypoint is a location in an image pinned down as a single point. It doesn't care what the target is — a joint, a finger knuckle, the corner of an object, anything.
- Each point carries a horizontal and a vertical position, and body-pose tools often output a confidence value for every point too. Instead of handing off the whole image, only that short bundle of numbers moves to the next step.
- In tools for poses and hands, which spots get marked, and in what order, is fixed ahead of time for each use. A hand commonly uses twenty-one spots; a whole body, seventeen.
- A point's real value sits less in the location itself than in the ratio of distances and the angle of the bends between points. Dividing one distance by another is what lets judgments hold up even as the camera moves farther away.
- Points on a dense list that has hardened into a standard, like a face, are often called landmarks.
The analogy
Countless stars scatter across the night sky, yet a constellation is drawn from only six or seven of them. The faint ones get skipped; only the bright, clearly placed stars get kept as points, and connecting them in a set order gives the shape.
How many stars get picked, and in what order they connect, is fixed ahead of time for each constellation. People who use the same star chart use the same stars for a constellation, and the same numbered star connects to the same next one. That's what lets those people talk about the same night sky under the same name.
A keypoint is one of those chosen stars. A photo holds hundreds of thousands of brightness values, but only a handful of spots worth pinning down get kept as points. Drop the rest, and the shape still holds.
In detail
Only a few points survive instead of the whole picture
A single photo is a dense block of brightness values laid out row by row. Carrying that whole block around is heavy, and it drags along useless information like clothing color or background. Extracting keypoints means keeping only a handful of location values from that block and letting the rest go.
What's left is short. A result with twenty points comes out to roughly forty to sixty numbers total. With a result that short, the judgment that follows is very light. Point-finding models are often built light for devices too, so they can run dozens of times a second right on the device instead of sending anything to a server — which is why webcam demos run without stuttering.
The remaining points alone can support plenty of judgments. Whether an arm is bent, a hand is open, or an object is tilted all gets worked out from the distance and angle between points. There's no need to look back at the picture itself.
Which spots and in what order gets fixed in advance
In tools for poses and hands, keypoints don't land just anywhere. A tool built for hands has a fixed list of spots — one wrist plus twenty finger joints. A tool built for the whole body has seventeen, covering shoulders, elbows, and more.
A fixed list means a fixed order too. Which numbered point is the thumb tip is always the same, so any program reading the result can just grab what it needs by number. Tools sharing the same list can swap results between them easily.
Different lists can't be mixed, though. Feed a twenty-five-point result into a classifier trained on seventeen points and the numbering misaligns, producing nonsense. That's why checking the point count comes first when switching tools. That said, some keypoints, like the feature points used to match photos, come without any list, placed wherever corners or patterns stand out.
A confidence value sometimes comes along too
Tools for body pose give every point a number for how likely that location is correct. Hand or face tools sometimes attach just one such value to the whole target. A clearly visible spot scores high; a hidden or blurry one scores low.
In tools that attach this value to each point, filtering on it is common in real services. Drawing every low-confidence point straight to the screen makes points look like they're shaking in midair. The usual fix is setting a threshold, hiding anything below it, or briefly carrying over the previous frame's position instead.
A high value doesn't always mean it's correct, either. A shadow or a pattern that resembles a joint can get a high score in the wrong place. The safer approach checks not just one point but whether its distance to neighboring points falls within what a real body could produce.
How it differs from landmark
The two words get used interchangeably a lot, but they carry a slightly different weight. Keypoint is a general term for any point that pins down a location. A joint, a finger knuckle, a cup's handle — mark a spot with a single point and it's a keypoint.
Landmark usually refers to a reference point where the target is fixed and the points are densely numbered into a settled standard. A face is the classic case. Point so-and-so is always the outer corner of the left eye, point so-and-so always the corner of the mouth — location and number are paired, so the same number means the same spot in any photo.
Put together, landmark is a branch of keypoint. The tightly settled, fixed-list branch is the one commonly called landmark, and the broader term one level up is keypoint. Still, some tools call hand or body points landmarks too, so the two words often get mixed.
More precisely
A single keypoint is expressed as a horizontal and a vertical position, and depending on the tool a confidence value or a front-to-back distance gets added. 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 straight at a location, the usual computation divides the image into a grid, fills each cell with a value for how likely that cell is the target spot, and picks the cell with the highest value. Build one such map per point, plus separate maps that carry the direction linking one point to the next, and it's possible to solve which points connect to which even when several people overlap in a scene.
The analogy breaks down somewhere too. The stars in a constellation are physically up there — someone just picks a few. A keypoint isn't marked on the image at all. It's a location guessed into being from countless hand-marked examples during training, an estimate of "probably around here." A star stays where it is even when faint, but a hidden keypoint can get placed somewhere that has nothing there at all.
Try it yourself
Common misconceptions
It's easy to think keypoint is a term only used for faces, but actually it covers any case where a location gets marked with a point — hands, bodies, objects, even maps.
It's easy to think more points is better, 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 keypoint and landmark are completely different things, but actually landmark is a branch of keypoint — it's mostly just the name for the branch where the list has densely settled, and the two words often get mixed.
One-line summary
A keypoint keeps only the locations in an image worth pinning down as points, and the distance and angle between those points become the raw material for judgment.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-17