A while ago I visited Shakespear park in Auckland and I took some photo scans of parts of a sandstone cliff. I developed an awesome workflow for making these high-quality assets that only uses Reality Capture and significantly reduces the time it takes to make assets.
The first step was to capture the image dataset. I did this with my Sony alpha NEX-6 which takes 16.1 Mpx photos that are nice and sharp. I took a total of 49 photos which I later pruned down to 43 good ones. I took all the photos in RAW as Reality Capture supports using them directly, and it should give better results.
For this asset, I developed a new workflow to create very high detail assets with normal maps using only Reality Capture. This let me use a very high-quality reconstruction (52M poly) to create detailed normal maps which I couldn’t do before because any other program dies even trying to import that much geometry. This is much quicker and easier than my previous asset workflow which involved importing and exporting files between 3 different programs and was very inflexible.
For alignment, I generally crank up all the settings to more reliably align all my photos.
Master mesh generation
To achieve the best quality normal maps, I used the 'High Detail' meshing option which took about an hour to complete and produced a model with 52.3M triangles. This model is too high-res to actually view in the software, but here is a render of it.
I then decimated the mesh to a workable quality of 10M triangles, cut off the blobby bits and textured the mesh.
Final mesh creation
For the final asset, I further decimate the master mesh and bake the normals of the extreme poly mesh (52.3M) to the final asset for a very detailed normal map. The following example is with the 50k poly standard-quality mesh shown in the Sketchfab viewer, but the process is the same for the 500k mesh.