Abstract | ||
---|---|---|
We propose a higher order conditional random field built over a graph of superpixels for partitioning natural images into coherent segments. Our model operates at both superpixel and segment levels and includes potentials that capture similarity, proximity, curvilinear continuity and familiar configuration. For a given image, these potentials enforce consistency and regularity of labellings. The optimal one should maximally satisfy local, pairwise and global constraints imposed respectively by the learned association, interaction and higher order potentials. Experiments on a variety of natural images show that integration of higher order potentials qualitatively and quantitatively improves results and leads to more coherent and regular segments. |
Year | DOI | Venue |
---|---|---|
2011 | 10.1007/978-3-642-24028-7_22 | ISVC (1) |
Keywords | Field | DocType |
conditional random field,higher order potential,higher order,embedding gestalt law,coherent segment,curvilinear continuity,global constraint,familiar configuration,capture similarity,higher order potentials qualitatively,natural image,image segmentation | Conditional random field,Computer vision,Graph,Pairwise comparison,Embedding,Pattern recognition,Computer science,Gestalt psychology,Image segmentation,Artificial intelligence,Curvilinear coordinates | Conference |
Citations | PageRank | References |
0 | 0.34 | 22 |
Authors | ||
3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Olfa Besbes | 1 | 22 | 3.38 |
Nozha Boujemaa | 2 | 1231 | 96.30 |
Ziad Belhadj | 3 | 125 | 10.56 |