Overlap, Tie Points, and Scale

Photogrammetry needs the scene to repeat across images.

Not because repetition looks nice, but because reconstruction software needs multiple observations of the same physical features to estimate geometry.

That is what image overlap is buying you.

Overlap creates shared evidence

Imagine a mapping strip:

Image 1: [ A B C D ]
Image 2:     [ C D E F ]
Image 3:         [ E F G H ]

Features C and D connect Images 1 and 2.

Features E and F connect Images 2 and 3.

Now the three images belong to one connected observation network.

If Image 2 were missing, the reconstruction might split into isolated sections.

That is why insufficient overlap can create gaps or partial reconstructions.

Front overlap and side overlap solve different coverage problems

In a grid-style aerial capture, neighboring images overlap:

• along the direction of travel;
• between adjacent flight lines.

Different mapping workflows recommend different amounts based on terrain, scene complexity, camera geometry, altitude, and desired product.

OpenDroneMap documentation, for example, publishes higher-overlap recommendations for many mapping and 3D workflows, especially around complex buildings and vegetation.

Treat those as workflow guidance—not a universal law for every camera and mission.

Your engineering responsibility is to justify the overlap needed for your supplied dataset and product goal.

More overlap is not automatically free

Higher overlap can give the software more repeated features.

It can also create:

• more images;
• more storage;
• longer processing time;
• greater mission duration;
• more battery demand;
• more chances for lighting or scene conditions to change.

So overlap is a tradeoff.

Too little can weaken reconstruction. Excess capture can increase workload without improving the product enough to justify the cost.

Tie points need texture

Two images can overlap physically and still match poorly.

Consider a featureless white roof.

The same large blank region may appear in several images, but there are few distinctive points to match.

Now compare a roof with:

• corners;
• vents;
• seams;
• painted marks;
• edges;
• texture changes.

The second scene gives feature-matching algorithms more unique structure.

Overlap tells the software where observations repeat. Scene texture determines whether those repeated observations are distinctive enough to match reliably.

Scale is a separate problem

Suppose your reconstructed point cloud has the correct shape but no trustworthy real-world scale.

The software might know that one wall is twice as long as another without knowing whether the first wall is 2 m, 20 m, or 200 m.

Real-world scale can be constrained using known reference information.

One method is ground control points: identifiable ground locations with known coordinates measured independently.

OpenDroneMap documentation describes GCPs as ground position measurements used for georeferencing and correcting distortion in mapping data.

The quality of the GCP coordinates and their distribution matter.

Ground control is not decorative markers

A target becomes useful control only when:

1. its real-world position is known to appropriate accuracy;
2. it can be identified clearly in the images;
3. it is distributed in a geometry useful to the project;
4. the processing workflow actually uses it correctly.

A bright checkerboard with made-up coordinates adds color, not accuracy.

Worked geometry: strong overlap, weak scale

A fictional image set reconstructs cleanly.

The software finds thousands of matches and produces a smooth 3D roof model.

But the dataset has:

• consumer GNSS image positions with unknown error;
• no ground control;
• no verified scale object;
• no independent checkpoint.

Can you claim the roof dimension is accurate to 2 cm?

No.

The strong relative geometry supports statements about shape and internal consistency better than it supports a tight absolute-accuracy claim.

The missing control evidence limits the measurement claim.

Relative versus absolute accuracy

This distinction is useful.

Relative accuracy

How well reconstructed features relate to one another inside the model.

Absolute accuracy

How well the model aligns with the real-world coordinate system or independently known positions.

A model can be relatively consistent while shifted or distorted in absolute coordinates.

That is why high-precision work needs independent reference evidence.

Checkpoints are different from control points

If every known reference point is used to force the model into alignment, you still need a way to test the result independently.

A checkpoint is a known point held out from the adjustment and used to evaluate the resulting model.

Conceptually:

control points → help fit/georeference model
checkpoints    → test the resulting model independently

This is the same idea as training versus testing in other technical fields: do not grade the model only on the evidence used to fit it.

Build an overlap-and-control plan

Use a supplied site rectangle and camera footprint diagram.

Create:

• proposed image rows;
• neighboring image footprints;
• overlap regions;
• one weak-coverage area;
• four fictional control/reference points;
• one independent checkpoint.

Then justify:

1. where feature matching is strongest;
2. where geometry may be weak;
3. why the control points are distributed where they are;
4. why the checkpoint is kept independent;
5. what product claim this network can support.

Misconception: GPS tags make every map survey-grade

Image geolocation is useful evidence.

Its accuracy depends on the positioning system, aircraft, conditions, metadata, processing, and workflow.

A coordinate existing in EXIF metadata does not prove the coordinate is accurate enough for a high-precision claim.

Measurement quality has to be demonstrated, not assumed from file format.

The model to keep

Photogrammetry quality depends on a connected observation network:

overlap + distinctive tie points + camera geometry + reference/scale evidence + independent validation.

A pretty orthomosaic can hide a weak network. Your job is to inspect the evidence underneath the pixels.