Image Set Quality Review

A photogrammetry dataset should be reviewed before you trust the reconstruction.

The best time to discover a coverage gap is before the original capture opportunity disappears.

This lesson uses a supplied image-set report rather than a live drone mission.

The fictional dataset

A roof-mapping set contains:

112 images
3 flight lines
camera mode: consistent
12 images flagged for blur
4 images show strong glare
1 flight-line turn created an overlap gap
3 ground control targets visible
1 independent checkpoint visible

The question is not:

Can the software process 112 images?

The question is:

After quality filtering, does the remaining observation network still support the intended product?

Review in layers

Layer 1 — File integrity and consistency

Check:

• files open correctly;
• expected image dimensions/mode are consistent;
• timestamps/metadata are plausible for the supplied project;
• no accidental unrelated images are mixed into the set.

Layer 2 — Image quality

Check:

• blur;
• exposure/clipping;
• obstruction;
• scene motion;
• unusual color/processing changes that could affect matching.

Layer 3 — Geometric coverage

Check:

• neighboring images overlap;
• adjacent flight lines connect;
• edges/corners of the project have coverage;
• complex vertical surfaces have appropriate viewpoints for the intended product.

Layer 4 — Reference evidence

Check:

• control points are identifiable;
• labels are correct;
• distribution covers the project reasonably;
• an independent checkpoint exists if the accuracy claim requires validation.

Only after these layers do you have a defensible input set.

Coverage maps expose hidden weakness

A list of filenames does not show geometry well.

Plot or sketch image footprints.

You may discover:

• one section covered by ten images;
• another section covered by only two;
• a diagonal gap between flight lines;
• an edge that appears in only one view.

The total image count can look healthy while one local region is fragile.

Photogrammetry quality is spatial.

Worked decision: reject 12 blurred images?

The 12 flagged frames are not equally important.

Suppose:

• eight have strong neighboring coverage;
• three are near the project edge with moderate redundancy;
• one is the only usable bridge across the turn gap.

The quality rule says the last image is too blurred for reliable use.

Do you keep it because geometry needs it?

No—not as if it became good evidence.

You document that rejecting it exposes a capture gap.

The correct project response may be recapture that area or reduce the product claim.

Keeping known-bad data only to make the software complete a model hides the real defect.

Use quality states, not one pass/fail bucket

Classify frames as:

• keep — meets current criteria;
• review — concern exists but image may remain useful;
• reject — fails defined criterion;
• replacement required — rejection creates unacceptable coverage loss.

That fourth state links image quality back to mission planning.

Ground-control visibility is an image-quality issue too

A control point coordinate may be excellent, but if the target center cannot be identified clearly in enough images, the control evidence becomes weaker.

Review:

• target sharpness;
• target size in pixels;
• lighting/contrast;
• obstruction;
• whether the same marker can be located consistently across views.

The coordinate and the imagery have to meet at one identifiable point.

Build a dataset quality dashboard

Use the supplied thumbnails/footprints and create:

Then state the dataset disposition:

• ready for processing;
• ready with stated limitations;
• partial processing only;
• recapture required.

Predict the processing failure before processing

Based on the input review, predict one likely output defect.

Examples:

• low-overlap corner may not reconstruct;
• moving cars may create ghosted geometry;
• glare may reduce matches on a roof section;
• weak control distribution may leave greater distortion far from control;
• blurred vertical views may weaken building-edge geometry.

Later, compare your prediction with the actual supplied processing report.

That is how you learn which input signals matter.

Misconception: failed reconstruction means bad software

Sometimes processing fails because the input evidence is insufficient.

A good quality review asks whether the software received a connected, sharp, referenceable observation network before blaming the algorithm.

Software settings matter, but the input set creates the ceiling.

Week 14 conclusion

Image-set QA is an engineering gate between capture and reconstruction.

You should be able to explain which images are trustworthy, where the network is geometrically weak, whether reference evidence is usable, and whether the dataset deserves processing or recapture.