Abstract
Volume parameterizations abound in recent literature, from the classic voxelgrid to the implicit neural representation and everything in between. Whileimplicit representations have shown impressive capacity and better memoryefficiency compared to voxel grids, to date they require training via nonconvexoptimization. This nonconvex training process can be slow to converge andsensitive to initialization and hyperparameter choices that affect the finalconverged result. We introduce a family of models, GA-Planes, that is the firstclass of implicit neural volume representations that can be trained by convexoptimization. GA-Planes models include any combination of features stored intensor basis elements, followed by a neural feature decoder. They generalizemany existing representations and can be adapted for convex, semiconvex, ornonconvex training as needed for different inverse problems. In the 2D setting,we prove that GA-Planes is equivalent to a low-rank plus low-resolution matrixfactorization; we show that this approximation outperforms the classic low-rankplus sparse decomposition for fitting a natural image. In 3D, we demonstrateGA-Planes' competitive performance in terms of expressiveness, model size, andoptimizability across three volume fitting tasks: radiance fieldreconstruction, 3D segmentation, and video segmentation.