Object Segmentation from Multi-Views Images with a Few Interactions

With the help of human-computer interaction(HCI), a monocular image can be segmented accurately. However, manual workload of HCI is remarkably heavy if a number of images are pending processing. Furthermore, it is difficult to construct a unified model that is universally applicable to different images. We propose an accurate segmentation algorithm for multi-view serial images with less HCI work. HCI is performed only on one serial image, and then the objects in the foreground of the other images can be segmented automatically. Based on the homography of the feature points built with a graphic model, the feature points in the HCI image and their corresponding points in other serial images are accordingly linked. The shortest path algorithm classifies the feature points into foreground or background, and searches the correspondence between the local features of serial images. Finally, the classified feature points are used as a priori knowledge of geodesic-star-convexity segmentation algorithm to accurately segment the serial images together. The comparison of hit rate and false alarm rate con-firms the accuracy and coherence of the proposed segmentation algorithm. Moreover, the local correspondence between the serial images can be used for multi-object segmentation.