REVISED DEPTH MAP ESTIMATION FOR MULTI-VIEW STEREO

Optical flow estimation is one of the popular methods to obtain the depth maps in multi-view stereo due to its high accuracy and robustness. In traditional optical flow estimation, the energy function contains three assumptions: intensity constancy assumption, gradient constancy assumption, and global smoothness assumption. In this work, we propose a local smoothness assumption to constrain the optical flow disparity in neighboring pixels. We first study the new smoothness term and its corresponding energy function, and present a practical iteration approach to minimize the energy function. Later we apply this new estimation method to the multi-view stereo system and obtain the depth maps of different image pairs. Our results demonstrate the good performance of the algorithm in acquiring smoothing surface when comparing to the traditional methods.