Interactive image segmentation

Abstract

The state-of-the-art single-image interactive segmentation algorithms are sensitive to the user inputs and often not able to produce accurate cutting contour with one-shot user input. They frequently rely on laborious user editing to refine the segmentation boundary. In the first part of this thesis, we propose a robust and accurate interactive image segmentation method based on the recently developed continuous-domain convex active contour model. The proposed method exhibits many desired properties for a good interactive image segmentation tool, including the fast segmentation speed, the robustness to user inputs and different initializations and the ability to produce a smooth and accurate boundary contour. Experimental results on a benchmark data set show that the proposed tool is highly effective and significantly outperforms the state-of-the-art interactive image segmentation algorithms. In the second part of this thesis, we extend our study to interactive segmentation of multi-view images, i.e., segmenting object regions from a sequence of calibrated images, which are taken on the same object from different viewing angles. In the case of a large number of images in the sequence, segmenting each image separately using interactive image segmentation techniques is time-consuming and requires a lot of user interaction and effort. The proposed method combines 2D interactive image segmentation and 3D object segmentation to exploit both the consistency constraint among different images in the sequence as well as the local information in each individual images to segment the sequence with a small amount of user interaction. Furthermore, we also introduce an editing tool for easily and arbitrarily refining the segmentation result. Experimental results show that the proposed method produces good segmentation result even with challenging image sequences and the editing tool is effective in refining the segmentation result.