Foreground Object Segmentation from Dense Multi-view Images

In order to extract foreground objects from dense multi-view images precisely and automatically, a level set evolution segmentation method without user interaction is proposed. Firstly, we make a statistical analysis of the straight lines in Epipolar Plane Image (EPI) and the EPI-lines corresponding to the foreground object are converted into original image space to get an initial contour. Then, we design a contour growing algorithm to shorten the gaps between broken edge segments and a morphological operation is utilized to obtain a closed exterior contour. Finally, a level set evolution without re-initialization is applied to drive the contour close to real object boundaries. Experimental results show that, our method can extract foreground objects from natural images more accurate and more effective than some user-assisted segmentation methods.