HOW PANORAMA STITCHING WORKS

Pair the points,
join the view.

A wide panorama is usually several ordinary photos.

A wide mountain lake view at sunrise with a photographer seen from behind turning a camera on a tripod — AI-generated image, not a photograph of a real person or place.

Find shared points
Throw out bad matches
Bend, then blend
Many photos, one picture

Stitching software finds distinctive points that appear in two overlapping pictures, works out how one picture must bend to fit the other, then blends the seam. It only works well when the photos give it enough shared detail.

Updated 16 October 2026

THE SHORT ANSWER

Match, check,
then blend.

Stitching has three main parts:
registration, calibration and blending.

MATCHING

Thousands of shared points

Feature detectors pick out corners and blobs that are repeatable and distinct, so the same spot can be found again in the next photo.

See the numbers

CHECKING

Throwing out bad pairs

A method called RANSAC repeatedly tries models on random sets of matched points and keeps the one that fits the most matches.

How to shoot

BLENDING

Hiding the seams

Light can differ between frames, and things move. Blending has to cope with parallax, lens distortion, scene motion and exposure differences.

What spoils it

A QUICK SHOT CHECK

What could spoil
the stitch?

Pick a situation to see why it helps or hurts. The bar shows how much trouble it causes the software.

Based on encyclopedia articles on image stitching and AutoStitch, and the 2004 SIFT paper. General information.

Stitch check

Choose a situationPlenty of overlap, steady exposureExposure changes between shotsPeople and cars movingPlain sky, little detailStepping sideways between shots

Plenty of overlap, steady exposureIdeal

The ideal set has a reasonable overlap, at least 15 to 30%, and consistent exposure between frames to reduce the chance of seams.

Exposure changes between shotsVisible seams

Since the light in two views cannot be guaranteed to match, stitching can leave a visible seam where brightness differs.

People and cars movingMismatches

Scene motion is one of the main issues stitching has to deal with, because the content of the overlap changes between frames.

Plain sky, little detailToo few points

Overlap is needed to have enough detectable features. A blank area gives the software little to match.

Stepping sideways between shotsParallax

Automatic stitchers such as AutoStitch require all the photos to be taken from a single point; moving the camera adds parallax.

Sources: Wikipedia, Image stitching and AutoStitch articles; Lowe, Distinctive Image Features from Scale-Invariant Keypoints (2004).

TWO NUMBERS WORTH KNOWING

Lots of points,
strict matching.

One figure on how many features a photo yields,
one on how false matches are weeded out.

≈ 2,000

Stable features in a typical 500 × 500 image

The paper that introduced SIFT features says a typical 500 by 500 pixel image gives about 2,000 stable features, depending on content and settings.

FEATURES · LOWE 2004

90%

False matches removed by a simple ratio test

Rejecting matches whose closest-to-second-closest distance ratio is above 0.8 removed 90% of false matches while discarding less than 5% of correct ones.

MATCHING · LOWE 2004

The SIFT features are invariant to image scale and rotation, which is why a zoomed or slightly turned photo can still be paired with its neighbour.

A busy town square with people walking across it, seen from slightly above in soft daylight — AI-generated image, not a photograph of a real person or place.

FOUR HABITS FOR A CLEAN STITCH

Give the software
something to pair.

These follow from what stitching software needs, whatever app or camera mode you use.

Overlap each frame
Aim for at least 15 to 30% overlap so neighbouring shots share enough detail.
Keep exposure steady
Consistent exposure between frames reduces the chance of visible seams.
Turn, don't walk
Take every photo from a single point, rotating the camera, to avoid parallax.
Wait for movement to pass
Moving people or vehicles in the overlap make frames disagree, so shoot when the scene is still.

HOW THE MATCH IS MADE

From a robot's eyes
to your phone.

One of the first interest point detectors was developed in 1977 for research on guiding a robot through a cluttered environment. Later detectors improved on it, and the 2004 SIFT paper described features that stay the same when an image is scaled or rotated.

Once points are matched, the software estimates a projective transform for each image, a change of coordinates that keeps straight lines straight, and places every photo on one surface before blending the overlaps.

A camera on a tripod with a panoramic head, seen close up on a rocky viewpoint above a valley — AI-generated image, not a photograph of a real person or place.

Questions

Questions people ask

How does a phone or computer stitch a panorama?

It finds distinctive points in overlapping photos, matches them between pairs of images, estimates how each image must be transformed to line up, and blends the overlaps into one picture.

How much should panorama photos overlap?

The Wikipedia article on image stitching says the ideal set has a reasonable overlap of at least 15 to 30%, with consistent exposure between frames.

Why does my panorama have a visible seam?

Light can differ between the two views, or the background can change. Parallax, lens distortion, scene motion and exposure differences all make seams more likely.

What is RANSAC?

RANdom SAmple Consensus, an iterative method that fits models to random samples of matched points and keeps the one that agrees with the most matches, ignoring outliers.

Why should I not move between shots?

Automatic stitching tools such as AutoStitch require all photos to be taken from a single point. Moving the camera sideways adds parallax, which blending struggles to hide.

Pair the points,
join the view.