Abstract: The segmentation of foreground objects in an image is useful in many applications. Traditional methods generally get the initial area of the foreground objects with the help of human interactions. For dataset with multiple images, the manual interaction is very tedious. To overcome this problem, an automatic method based on consistency analysis is proposed to segment foreground objects from multiple images. First, the conversion between different views is achieved with the help of the prior knowledge of the three dimensional scene. Then initial label-ling results of foreground and background are obtained by consistency analysis on each pixel. Finally, an energy equation is constructed based on the initial labelling results and is optimized iteratively and continuously to obtain the accurate contour of foreground objects. The experimental results show that the proposed method can extract foreground objects accurately and the segmentation results can be used for 3D reconstruction.