Photogrammetry glossary: the terms, in plain words

47 of 47 terms.

ABCDEFGILMNOPRSTV

Absolute accuracy

How close a point in the model is to its true position on the Earth, in a named coordinate system. Needs ground control or RTK/PPK; without them a model can be perfectly shaped and metres out of place.See Relative accuracy, Ground control point.

Aerial triangulation

The classic name for solving the positions and orientations of all the photographs together from tie points and control, before anything else is made. Bundle adjustment is how it is done now.

Bundle adjustment

The single big calculation at the heart of photogrammetry: adjust every camera position, orientation, lens parameter and 3D tie point at once until the reprojection error in every photograph is as small as it can be. 'Bundle' is the bundle of light rays through each camera.See Tie point, Reprojection error.

Check point

A surveyed point that was not used to control the model, kept back to measure the model's error afterwards. Accuracy claims are only honest when made against check points.

Close-range photogrammetry

Photogrammetry with the camera within a few metres or tens of metres of the subject: objects, machinery, building faces, accident scenes. Scale usually comes from scale bars or coded targets rather than GNSS.

Coded target

A printed pattern (a ring of black and white segments, or a checkerboard) the software recognises automatically and locates to a fraction of a pixel. Used as ground control points and for scale in close-range work.

Dense cloud

The point cloud made after the cameras are solved, by finding depth at nearly every pixel: millions of coloured points. The mesh, orthomosaic and elevation models are built from it.See Sparse cloud, Point cloud.

Digital elevation model (DEM)

Any grid of heights; used loosely for both a surface model and a terrain model, so check which one a file is before measuring on it.See Digital surface model (DSM), Digital terrain model (DTM).

Digital surface model (DSM)

A grid of the height of the first thing seen from above: roofs, canopy, stockpiles, trucks, and bare ground where it is bare. What stockpile volumes are measured on.

Digital terrain model (DTM)

A grid of the height of the ground alone, with vegetation and objects removed by classification or filling. What cut and fill, contours and grading are measured on.

Direct georeferencing

Positioning the model from the recorded position of each photograph (RTK or PPK) rather than from targets on the ground. Good horizontally; vertically it needs oblique photos or a few targets to hold.

Doming (bowl effect)

A systematic bend in a model made from parallel, nadir-only flight lines at one height: the middle comes out high or low against the edges because a small lens error and a curved world look alike to the solver. Obliques, cross lines and ground control break it.

EPSG code

A number naming a coordinate reference system, from a registry kept by the oil-industry geodesists who started it. 32632 is UTM zone 32N on WGS 84; 4326 is plain latitude and longitude.

Exterior orientation

Where a photograph was taken from and which way the camera pointed: three positions and three angles for every picture.

Feature (keypoint)

A small, distinctive patch in a photograph that can be recognised again from another angle, described by the pattern of light around it. Thousands per picture; the raw material for tie points.

Georeferencing

Giving a model real-world coordinates, by ground control points, by the positions of the photographs, or both.

Global shutter

A sensor that exposes every pixel at the same instant, so a moving camera does not distort the geometry. The alternative, a rolling shutter, reads rows in sequence.See Rolling shutter.

GNSS

Global navigation satellite systems: GPS, GLONASS, Galileo, BeiDou and the rest, together. A plain receiver positions to a metre or two, sometimes more; RTK and PPK bring that to centimetres.

Ground control point (GCP)

A target on the ground whose position has been surveyed, placed so that it appears in the photographs. The model is pinned to it. The single most effective thing for vertical accuracy.

Ground sample distance (GSD)

The size of one pixel on the ground, usually in centimetres: sensor pixel size × height ÷ focal length. Nothing in the survey can be finer than it, and accuracy is usually quoted as a multiple of it.

Interior orientation

The camera's own geometry: focal length, principal point and lens distortion. Solved in calibration or self-calibrated in the bundle adjustment.

LiDAR

Light detection and ranging: measuring distance by timing laser pulses. An active sensor that works at night and sees through gaps in vegetation, where a camera does not.See Multiple returns.

Mesh

A surface of triangles joined over the point cloud, with the photographs projected onto it as texture. The 3D model you can orbit.

Multi-view stereo (MVS)

The dense matching step: with the cameras known, find the depth at each pixel by comparing many photographs along known lines of sight. Makes the dense cloud.

Multiple returns

One laser pulse reflecting more than once on its way down: first from the top of a canopy, then from branches, last from the ground. The last returns are what a terrain model under trees is built from.

Nadir

Straight down. A nadir photograph looks vertically at the ground; an oblique one is tilted.

Oblique photograph

A photograph taken with the camera tilted away from vertical. Shows walls and faces a nadir picture cannot, and stiffens the geometry of a block of photos against doming.

Orthomosaic

The photographs re-projected so that every pixel is seen straight down at a constant scale, in map coordinates, and blended into one image. Positions, lengths and areas can be measured on it; heights cannot.

Overlap

How much of each photograph is also in its neighbours: front (along the flight line) and side (between lines). 75% front and 65% side is the usual minimum; more over vegetation and buildings.

Parallax

The shift of a point between two photographs taken from different places. The further the point, the smaller the shift; measuring the shift is how depth is found.

Point cloud

A set of points in space, each with a position and usually a colour or intensity. What photogrammetry and laser scanning both produce before any surface is made.

PPK (post-processed kinematic)

Correcting the drone's recorded GNSS positions after the flight against a base station's log, to a few centimetres. The same result as RTK without needing a radio link during the flight.

Relative accuracy

How true the model is to itself: whether a 10 m wall measures 10 m, whatever the model's position on the Earth. Photogrammetry's relative accuracy is good even without control.

Reprojection error

After the solution, the distance in pixels between where a tie point was seen in a photograph and where the solved model says it should appear. The bundle adjustment minimises it; a few tenths of a pixel is good.

RMSE

Root mean square error: the usual single figure for accuracy, computed from the differences at the check points. Larger than the typical error, because big misses count for more.

Rolling shutter

A sensor read out a row at a time over a few tens of milliseconds. From a moving camera each row sees the ground from a different place and straight edges lean. Most small drone cameras have one.

RTK (real-time kinematic)

Correcting the drone's GNSS position live, during the flight, against a base station or a network, to a few centimetres. Each photograph then knows where it was taken to that accuracy.

Scale bar

A bar of known length placed in close-range photographs so the model has a true scale, since nothing in a photograph says how big anything is.

SLAM

Simultaneous localisation and mapping: a scanner that works out where it is from what it sees as it is carried along, with no GNSS. How handheld laser scanners map a building or a tunnel.

Sparse cloud

The tie points in 3D, solved in the bundle adjustment: tens of thousands of points that fix the cameras before the dense cloud is made.

Stereo pair

Two overlapping photographs taken from different positions, from which depth can be seen and measured: the classic unit of photogrammetry before structure from motion.

Structure from motion (SfM)

The modern method: from any number of overlapping photographs, find features, match them, and solve the cameras and the points together without knowing the camera positions beforehand. What drone photogrammetry software does.

Terrestrial laser scanner (TLS)

A laser scanner on a tripod that sweeps its surroundings and returns millions of points with millimetre accuracy to a few hundred metres. Buildings, plants, tunnels, quarry faces.

Texture

The photographs mapped onto a mesh so it looks like the real thing. Also, the fine detail on a surface that photogrammetry needs in order to match it: a blank wall has none.

Tie point

One spot on the ground seen in two or more photographs, with its pixel position in each. Tens of thousands of them tie the block of photographs together.

Time of flight

How LiDAR measures distance: time a pulse of light out and back, and multiply by the speed of light. Phase-shift scanners measure the same thing by comparing waveforms.

Volume (cut and fill)

The space between two surfaces: a stockpile over its base, or today's ground against a design. Computed from the elevation model, never from the orthomosaic.