Animated 3D Creatures from Single-view Video by Skeletal Sketching

Abstract

Extraction of deformable 3D geometry is not accessible to casual users, as it either requires dedicated hardware or vast manual effort. Inspired by the recent success of semi-automatic 3D reconstruction from a single image, we introduce a sketch-based extraction technique that allows a fast reconstruction of a dynamic articulated shape from a single video. We model the shape as a union of generalized cylinders deformed by an animation of their axes, representing the "limbs" of the articulated shape. The axes are acquired from strokes sketched by the user on top of a few key frames. Our method bypasses the meticulous effort required to establish dense correspondences when applying common structure from motion techniques for shape reconstruction. Instead, we produce a plausible shape from the fusion of silhouettes over multiple frames. Reconstruction is performed at interactive rates, allowing interaction and refinement until the desired quality is achieved.