A digital twin is a virtual copy of a place used to test robots. It needs realistic images and a surface model that tells simulated objects where they can bump into things. Ordinary reconstruction tools can lose the edges of fisheye images or fail to create that surface model.
GOFish uses fisheye photos and measurements of the camera lens to build a Gaussian-splat scene, rendered from many small, positioned translucent shapes. It also extracts an aligned mesh, a surface made of triangles, while accounting for lens shading and blur separately from the scene.
Across wide-angle tests, its peak signal-to-noise ratio, a measure of how closely rendered images match photos, beats the strongest prior method by 2.28 and 1.97 decibels. It also achieves top mesh-accuracy scores on reported tests, combining realistic views with collision geometry without Light Detection and Ranging scans.