3D Structure Refinement of Nonrigid Surfaces through Efficient Image Alignment

摘要: Given a template image with known 3D structure, we show how to refine the rough reconstruction of nonrigid surfaces from existing feature-based methods through efficient direct alignment. Under mild assumption that barycentric coordinates each point on surface keep constant, prove and input are correlated by piecewise homography, based which Lucas-Kanade alignment method is proposed iteratively recover an inextensible even poor texture sharp creases. To accelerate method, equivalent but much more as well, in most time-consuming part Hessian can be pre-computed result combining additive inverse compositional expressions. Sufficient experiments both synthetic real images demonstrate accuracy efficiency our methods.