Computer Vision News - March 2023

5 nerf2nerf with Lily Goli of appearance and geometry , ” she explains. “ We wanted this to work between any two NeRFs that have the same object in them, so with different lighting or at a different time of the day, the appearance of the object, the color of it, might change. A cup, for example, in daylight might be very white, but at night, it might be gray, so we can’t use the radiance colors outputted from NeRF directly to compare two points on the two NeRFs to see if they match. ” Instead, Lilyhad to extract a representation that showed the object’s geometry to compare it . Geometry is reliable because the possibility of the same instance of an object, the goal is to find the perfect rigid transformation to align these objects . However, this is more challenging as NeRFs are implicit representations without explicit access to the radiance, color, or other geometry used to compare images. Therefore, Lily and the team had to devise creative methods for performing this registration. “ Themain challenge here is that NeRFs bake in the appearance of the objects inside their geometry, so there is no clear decoupling