Logo image
Interactive Image Segmentation via Superpixel-Wise Label Propagation
Thesis

Interactive Image Segmentation via Superpixel-Wise Label Propagation

Lee, Yu-An
Masters, 國立清華大學, 資訊工程學系
2015

Abstract

互動式影像分割 超像素 標籤擴散 半監督式學習 Interactive Image Segmentation Superpixel Label Propagation Semi-Supervised Learning
The performance of an interactive image segmentation method can be evaluated not only by the label accuracy, but also the interaction time. Generally, the conventional approaches evaluate their execution time by a single-iteration test with a set of input labels. In practice, however, the user would input the labels multiple times until obtaining satisfactory results. In a case of the multi-iteration segmentation task, the interaction time could be divided into three parts. Firstly, the times of interactions. The total iterations of user interaction should be as less as possible. Secondly, the waiting time. The waiting time between two iterations should be short. Thirdly, the quantity of the given labels. Less information provided is the better. In this paper, we propose an interactive segmentation method based on over-segmentation and label propagation technique. We map the superpixels to a five dimensional feature space, then get the segmentation result by propagating the given labels. The proposed “Superpixel-Wise Label Propagation” (SLP) method can handle the object with mesh or long structures by just a few given information. The interaction time is also concerned in SLP. We adopt the segmentation result as the input labels of the next iteration. It decreases not only the computational time, but also the total times of interactions. Furthermore, SLP achieves state-of-the-art of GrabCut dataset comparing with other methods.

Metrics

1 Record Views

Details

Logo image