Continuous remeshing for inverse rendering

Abstract

AbstractWe present a novel method for joint optimization and remeshing and apply it to inverse rendering. Rapid advances in differentiable rendering during the last years paved the way for fast inverse rendering of complex scenes. But serious problems with gradient‐based optimization of triangle meshes remain. Applying gradient steps to the vertices can lead to mesh defects, such as flipped triangles, crumpled regions, and self‐intersections. Choosing a good vertex count is crucial for the optimization quality and performance but is usually done by hand. Moreover, meshes with fixed triangulation struggle to adapt to complex geometry. Our novel method tackles all these problems by applying an adaptive remeshing step in each single iteration of the optimization loop. By immediately collapsing suspicious triangles, we avoid and heal mesh defects. We use a closed‐loop‐controlled location‐dependent edge length. We compare our solution to state‐of‐the‐art methods and find that it is faster and more accurate. It produces finer meshes with fewer defects, requires less parameter tuning and can reconstruct more complex objects.