| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 14.76 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Markov Random Fields (MRFs) are a populartool in many computer vision problems and faithfully modela broad range of local dependencies. However, rooted in theHammersley-Clifford theorem, they face serious difficulties inenforcing the global coherence of the solutions without using toohigh order cliques that reduce the computational effectiveness ofthe inference phase. Having this problem in mind, we describea multi-layered (hierarchical) architecture for MRFs that isbased exclusively in pairwise connections and typically producesglobally coherent solutions, with 1) one layer working at the local(pixel) level, modelling the interactions between adjacent imagepatches; and 2) a complementary layer working at theobject(hypothesis) level pushing toward globally consistent solutions.During optimization, both layers interact into an equilibriumstate, that not only segments the data, but also classifies it.The proposed MRF architecture is particularly suitable forproblems that deal with biological data (e.g., biometrics), wherethe reasonability of the solutions can be objectively measured.As test case, we considered the problem of hair / facial hairsegmentation and labelling, which are soft biometric labels usefulfor human recognitionin-the-wild. We observed performancelevels close to the state-of-the-art at a much lower computationalcost, both in the segmentation and classification (labelling) tasks
Description
Keywords
Hair analysis Soft Biometrics Visual Surveillance Homeland Security