Surface normal integration for convex space-time multi-view reconstruction

We show that surface normal information allows to significantly improve the accuracy of a spatio-temporal multi-view reconstruction. On one hand, normal information can improve the quality of photometric matching scores. On the other hand, the same normal information can be employed to drive an adaptive anisotropic surface regularization process which better preserves fine details and elongated structures than its isotropic counterpart. We demonstrate how normal information can be used and estimated and explain crucial steps for an efficient implementation. Experiments on several challenging multi-view video data sets show clear improvements over state-of-the-art methods.