PPR: Physically Plausible Reconstruction from Monocular Videos

Given casually-captured monocular videos (left), PPR builds 3D models of articulated objects and the surrounding environment. Naive kinematic reconstruction (middle) generates a family of solutions, some containing inconsistent physical support and contact dynamics (blue and green color), such as floating or walking with sliding feet. We show that differentiable physics simulation acts as effective regularizer for improving the physical plausibility of visual reconstruction algorithms. As PPR reconstructs the dynamics scene, it also drives a ragdoll in a physics simulator to track the kinematic reconstruction. This ensures the reconstructions are statically stable with ground contact (right), and the center of mass is projected within the support polygon (marked with red). PPR also reports physics estimations, such as ground reaction forces (red arrows) and center of mass (green arrow).