無校正影片中多材質物體深度之重建

Abstract

本論文提出一個能從影像序列中還原物體表面演算法，以應用在三維顯示器中。針對單一方向光源環境，我們能還原多材質和非剛體物體表面。還原真實物體表面一直是很難的課題，shape-from-shading和multiple-view是最常使用的技術。但是shape-from-shading只能處理單一材質且表面連續的物體，multiple-view則無法保留較細微的表面，同時也需要耗時的配對對應點。因此， 我們結合 shape-from-shading 和 non-rigid structure-from-motion的技術來重建更細微完整的表面。

In this thesis, we propose a method to reconstruct real object in a video sequence which can be used for the 3D display. Under a single directional lighting condition, the target object can be multi-material、non-rigid. Shape recovery of real object from a video sequence is a difficult subject. Shape-from-shading and multiple-view technique are the most common methods. Shape-from-shading can only deal with single-material smooth objects. The multiple-view technique acquires the point clouds without the surface detail and needs time-consuming correspondence matching. For these reasons, we combine shape-from-shading and non-rigid structure-from-motion techniques to reconstruct more detailed surfaces.