A two-stage cascaded foreground seeds generation for parametric min-cuts

• Shao-Mei Li,
• Jun-Guang Zhu,
• Chao Gao,
• Chun-Wei Li,

Vol. 10, No. 11, November 29, 2016
10.3837/tiis.2016.11.020, Download Paper (Free):

• Object proposal
• foreground seed
• two-stage cascaded ranking
• parametric min-cuts

Abstract

Parametric min-cuts is an object proposal algorithm, which can be used for accurate image segmentation. In parametric min-cuts, foreground seeds generation plays an important role since the number and quality of foreground seeds have great effect on its efficiency and accuracy. To improve the performance of parametric min-cuts, this paper proposes a new framework for foreground seeds generation. First, to increase the odds of finding objects, saliency detection at multiple scales is used to generate a large set of diverse candidate seeds. Second, to further select good-quality seeds, a two-stage cascaded ranking classifier is used to filter and rank the candidates based on their appearance features. Experimental results show that parametric min-cuts using our seeding strategy can obtain a relative small pool of proposals with high accuracy.

Statistics

Cite this article