Segmenting natural images with the least effort as humans
Qiyang Zhao · 2015
Given that natural image segmentation is well-known as an ill-posed problem, then how can we design an algorithm to obtain good performance as human subjects? A choice is to learn from human segmentations as MCG [3] which obtains the state-of-the-art performance and fairly high computational efficiency. Then a question arises here: what should we learn? The way all existing literatures exploit human segmentations ignores a basic fact that human segmentations are produced by human operations. The human segmentation process would inevitably fulfill the human behavior’s principles including the least effort principle (LEP): a human will strive to solve his problem in such a way as to minimize the total work that he must expend [1]. The principle gives us a new insight into our problem. Suppose you are a human subject in the BSDS segmentation experiment, and are required to segment images into pieces under the ending instruction: each piece contains only one single distinguished thing. Then your total effort F of segmenting an image should including understanding images (U) to guide following operations, and tracing boundaries (T ) by hands with a mouse: F (I,S) =U (I)+T (S). Then from the viewpoint of LEP, human subjects would like to find an acceptable segmentation S by minimizing their effort, or formally,