Object Segmentation Using Structural Relationship between Super-pixels
Yonghui Gao, Lei Zhou, Xiaoxiao Li · 2016
We address the problem of describing and integrating long range information efficiently, such as the information demonstrated by super-pixels (patches), into conditional random field (CRF) model for object segmentation.For those purpose, a novel structural relationship between patches are defined for evaluating super-pixels' similarity.The structural relationship between super-pixels will focus on whether two patches can display similar information of objects' global appearance.Furthermore, a regression model is learned for super-pixels classification based on analyzing their structural relationship between super pixels and initial object hypothesis.Finally, a pixel-level CRF model that integrates information of color, texture and super-pixels is constructed to obtain segmentation results.Compared with traditional super-pixels or solely pixels based model, our method can combine the complementary information provided by pixels and super-pixels and generate better performance.