MARF: The Medial Atom Ray Field object representation

We propose Medial Atom Ray Fields (MARFs), a novel neural object representation that enables accurate differentiable surface rendering with single network evaluation per camera ray. Existing ray fields struggle multi-view consistency and representing discontinuities. MARFs address both using medial shape representation, dual of solid geometry yields cheap geometrically grounded normals, in turn enabling computing analytical curvature despite the having no second derivative. map to multiple intersection candidates, subject ray-sphere testing. illustrate how learned quantities applies sub-surface scattering, part segmentation, aid space articulated shapes. Able learn priors, may prove useful for tasks like retrieval completion, among others. Code data can be found at https://github.com/pbsds/MARF.