7AAVDH22 Chap.4 From Photographs to a Defensible 3D Model
From Photographs to a Defensible 3D Model
Follow the reconstruction pipeline
The modelling seminar introduces alignment, point clouds, meshes, textures and optimisation. Each stage changes the representation. Alignment estimates camera relationships from shared features. A point cloud contains reconstructed spatial samples. Meshing joins points into a surface. Texturing projects image appearance.
Optimisation changes geometry or texture so the model can perform in a target experience.
Judge each stage with the right test
A visually persuasive final model can conceal weak evidence. Alignment should be inspected for split camera groups, drift and orientation. Point clouds expose density, noise and missing regions before surfaces hide them. Mesh review identifies bridges, holes and excessive smoothing.
Texture review checks seams, blur, baked lighting and misprojection. Optimisation is evaluated against both performance and the heritage feature the user must inspect.
Trace errors backwards
When an anomaly appears, compare it with source photographs and direct observation. Move backwards from texture to mesh, points, alignment and images until the first unsupported transformation appears.
Classify the cause as occlusion, weak surface texture, reflection, motion, inconsistent light, poor overlap or a processing decision. Recapture, delete false geometry, preserve uncertainty or narrow the use according to the evidence available.
Keep observed and generated surfaces distinct
Automatic hole filling and manual cleanup can create plausible geometry where photographs supplied none.
Such repairs may be useful for a display derivative, but they need documentation and should not be treated as observed heritage evidence. A false bridge over an opening should be removed when it contradicts source views. A low-confidence surface can remain if its status and prohibited uses are clear.
Optimise for a named encounter
A preservation master and delivery model serve different roles.
Geometry can be reduced where detail is not interpretatively important, while edges, inscriptions and diagnostic surface traces may require protection. Test on the intended device and connection.
A smooth frame rate is valuable, but an interaction fails if optimisation removes the evidence it was designed to reveal.
Make a checkable reconstruction claim
End with intended use, observed evidence, major transformation, remaining uncertainty and fitness for purpose. Replace accurate with the feature and comparison actually checked.
Technical consistency cannot settle cultural authenticity, social authority or embodied meaning, but transparent reconstruction gives later analysis a defensible foundation.
What this chapter covers
- 01
Camera Alignment
- 02
Point Cloud
- 03
Mesh Topology
- 04
Texture Projection
- 05
Model Optimisation
- 06
Fitness For Purpose
Worked example · free
Remove an invented surface
- 2Compare the anomaly with source views.
- 3Trace the cause through mesh, points and images.
- 4Choose repair or recapture and document limits.
Key terms
- Camera Alignment
- Camera Alignment estimates the relative positions of photographs from stable shared visual features.
- Point Cloud
- Point Cloud is a set of reconstructed spatial samples whose density and noise reveal evidence quality.
- Mesh Topology
- Mesh Topology describes how vertices, edges and faces connect to form the reconstructed surface.
- Texture Projection
- Texture Projection maps image appearance onto geometry and may retain seams, lighting and capture artefacts.
- Model Optimisation
- Model Optimisation changes geometry or textures to meet delivery constraints while attempting to preserve essential evidence.
From Photographs to a Defensible 3D Model FAQ
Why can a realistic model still be wrong?
Texture can conceal distorted or generated geometry, and visual polish does not show whether views aligned correctly. Compare silhouette, scale, diagnostic features and uncertain regions against source photographs and paradata before making fidelity claims.
Should holes in a mesh always be filled?
No. Fill only when the action is justified for the intended use and the generated surface is documented. If the hole represents missing observation or contradicts source views, preserving or exposing uncertainty is more defensible than creating a plausible completion.
How should optimisation be evaluated?
Test the delivery model on the target device and connection while performing the essential interpretative task. Report what improved, what detail changed and whether that loss affects the heritage purpose. Preserve a richer master separately when possible.
Assessment move
Draw the pipeline from photographs to alignment, point cloud, mesh, texture and delivery model. For each arrow, name one transformation and one possible failure. Practise tracing an invented anomaly backwards until its cause becomes visible. Then write separate fitness statements for a web orientation model and a measurement model.
If the same statement fits both, the intended use and evidence threshold are not yet specific enough. Take a polished model screenshot and list three problems the image could conceal. Explain which pipeline stage would expose each problem and which source comparison would be required.
Finish by separating master preservation needs from delivery performance needs, then state why optimisation can improve access while weakening later inspection if the relationship between files is lost. Retain one diagnostic view that makes the most serious uncertainty visible.