I’ll paste in from a post begun over the RealityCapture forum:
As a favor to a colleague, I scanned a prosthetic hand, can share what went into capture and QC, but to make a long story short, I was shocked to see how poorly the model turned out using High (see RC_Hand1.jpg). My friend then used his iPhone 11 to produce decent model using Polycam. Given his photography was suboptimal, lots of soft focus, poor lighting, motion blur, and given my 38 MP photography was well lit, sharp pixels, how to account for these two shocking twists - lesser camera/lighting and RC performing behind Polycam? (See RC_Hand2.jpg for RC’s processing of my colleague’s photography)
Granted with my 21mm my cameras worked at greater distance to subject, the finest details in SOME of his pics appears greater. All the same, that doesn’t explain the results.
About the material, it’s a rubber, fairly homogenous color, so we scuffed it up with sand paper, added some powdered chalk coloring, have the impression there’s a fair amount of translucency in play, my on-camera flash possibly exacerbating the subscatter. Even if that explain why his photos did better than mine in RC, we’re left with his photos performing better in Polycam than in RC.
I happily share both datasets. Thanks for insights.
So, in the meanwhile, Ondrej worked with me to gradually unravel the mystery, first having me change my Alignment settings, Max error had been the default 2.0, brought down to 0.45, also cranking detector sensitivity from Med to Ultra. while that brought my max, median, and mean errors down significantly, I only spotted minor improvements in the mesh, the fingers being the main issue, but the surface topology along the wrist and forearm appearing super noisy. I had suspected the rubber material might have invited major subsurface scatter from my strobe (my friend’s iPhone 11 was only ambient lit), but no, keep reading. When I uploaded the project for Ondrej to run on his end, his hand appeared close enough to the screen capture I saw of the Polycam reconstruction, the texture possibly hiding lots of artefacts of course. But why with the same settings were my 60 images producing a far improved reconstruction of 57 cameras, where mine with but 56 was so highly flawed. I had cleared cache, new project, tried working with XMPs from his alignment, minor improvements turned up, but not until a clean install (of the same version) did the reconstruction appear optimized, which I’ve attached a snip with all settings exposed.
Takeaway, clean install whenever upgrading RC, and upping the detector sensitivity/lowering max error can help.
Thanks, Ondrej, super responsive and helpful!
