Ambient lit photogrammetry might suffice for subject matter  bathed in adequate light, especially with applications not requiring relighting. But, who wants to forego relighting? In truth, ambient lit photogrammetry can be relit, and often is, but not without serious caveats.

• Shadows translate into noise
• Noise translates into ratty geometry, if not useless data
• Manual cleaning is expensive
• CG tricks used to fake roughness – expensive
• Realism – which wins out, nature or an artist’s interpretation?

So, what about  NeRFs? They’re ambient lit and look great. Some have argued photogrammetry is dead. The utility of polygonal geometry using PBR textures has its arguments, but one inescapable fact is that ambient lighting poses a serious constraint on capture, never mind where you take the data. Empirical AC/R or inferred PBR via AI? It’s a false dichotomy. AC/R and machine learning compliment one another. The former acts against “garbage in, garbage out” quality of data and provides a leg up to optimize post-processing. Armed with the best input, machine learning further automates post-processing and can make the strongest inferences about material reflectance properties.

Under ideal lighting conditions, outdoors with unobstructed diffuse light, common photogrammetry produces acceptable results. For this reason, most common photogrammetry limits itself to evenly lit subject matter out in the open, steering clear of highly occluded subject matter, where every protruding object/feature casts shadows upon every other object/feature. The resulting noise in the shadows translates into geometry riddled with artifacts, the worst of which remain hidden so long as common photogrammetry is rendered unlit. In order to relight common photogrammetry, users absorb costly manual labor required for cleanup. Alternately, they may augment ambient lighting during capture, which then introduces its own set of constraints and issues with problematic specular reflections.