A digital-to-physical-to-digital protocol for comparing 2D Gaussian Splatting, photogrammetry, NeRF and iPad Pro/Polycam RGB-D scanning on a 3D-printed object

Comparisons of 3D reconstruction pipelines depend on the reference against which they are scored. We describe a digital-to-physical-to-digital protocol in which a digital design is 3D printed, the print is captured, and every reconstruction is compared with the design that produced it. We applied it to a 19-cm printed sculpture, captured as smartphone video and with an iPad Pro, to compare 2D Gaussian Splatting (2DGS), photogrammetry, NeRF (nerfacto with NeRF2Mesh) and the iPad Pro/Polycam RGB-D pipeline. Geometry was scored by bidirectional distances to the design, with spatial block bootstrap, registration and fabrication analyses; appearance was scored on 30 views held out from every image-based pipeline, inside object masks derived from the photographs alone. Symmetric root-mean-square deviations from the design were 1.54 mm (2DGS), 1.80 mm (photogrammetry), 3.40 mm (iPad Pro/Polycam) and 5.31 mm (NeRF). The difference between 2DGS and photogrammetry was small relative to its uncertainty, although 2DGS was closer to the design over most of the surface. On the held-out views, object PSNR was 32.4, 26.3, 19.4 and 13.1 dB for 2DGS, NeRF, photogrammetry and iPad Pro/Polycam. For this object and configuration, 2DGS gave the lowest measured geometric deviation and the highest measured photometric fidelity.