Differentiable rendering using RGBXY derivatives and optimal transport
DescriptionWe present a novel differentiable rendering framework that is reformulated from the Lagrangian view. Inspired by fluid simulation, traditional differentiable rendering approach can be viewed as the Eulerian method due to the fact that image derivatives are computed locally on a fixed grid of screen-space pixels. In contrast, our method tackles this problem from the Lagrangian viewpoint, where pixels are considered as fluid particles that are movable across the image plane locations, capturing global and long-range object motions efficiently. We guide the underlying geometry point at each pixel through robust correspondence finding and track how they change with respect to the optimizable scene parameters. As the efficacy of our technique hinges with the quality of the searched correspondence, we present a study of different distance calculation metrics including bipartite graph matching, flow estimation and optimal transport. Further, we evaluate the effectiveness of our proposed framework on various inverse rendering applications and demonstrate superior convergence behavior compared to state-of-the-art baselines.
Event Type
Technical Papers
TimeTuesday, 6 December 202210:00am - 12:00pm KST
Registration Categories