Abstract
We propose an end-to-end inverse rendering pipeline called SupeRVol thatallows us to recover 3D shape and material parameters from a set of colorimages in a super-resolution manner. To this end, we represent both thebidirectional reflectance distribution function (BRDF) and the signed distancefunction (SDF) by multi-layer perceptrons. In order to obtain both the surfaceshape and its reflectance properties, we revert to a differentiable volumerenderer with a physically based illumination model that allows us to decouplereflectance and lighting. This physical model takes into account the effect ofthe camera's point spread function thereby enabling a reconstruction of shapeand material in a super-resolution quality. Experimental validation confirmsthat SupeRVol achieves state of the art performance in terms of inverserendering quality. It generates reconstructions that are sharper than theindividual input images, making this method ideally suited for 3D modeling fromlow-resolution imagery.