Semantic Object Segmentation - Video Segmentation and Its Applications - page 70

Fig. 3.7 Confusion matrix of object segmentation by TextonBoost [ 40 ] on the MSRC 21 database.
The figure is reproduced from [ 40 ]

The experimental evaluation showed that although the texture-layout potentials had
the most significant contribution to semantic object segmentation, CRF significantly
improved the accuracy of results.

3.3.2.3 Other Approaches Based on Conditional Random Fields

Other semantic object segmentation approaches based CRF were proposed.
Fulkerson et al. [ 42 ] treated superpixels [ 43 ], which were small regions obtained
from a conservative oversegmentation, as basic units of segmentation. They as-
sumed that superpixels allowed to measure histograms of visual words on a natural
adaptive domain rather than on a fixed patch window. Moreover, superpixels tended
to preserve boundaries and created more accurate segmentation. A one-vs-others
SVM classifier with a RBF-

2 kernel was constructed on the histograms of visual
words found in each superpixel. This local classifier was used in a CRF operating on
the superpixel graph. CRF was used to add spatial regularization by requiring that if
two neighboring superpixels share a long boundary and were similar in appearance,
they tended to have the same class label. It discouraged small isolated regions and
reduced misclassifications that occurred near the edges of objects. He et al. [ 44 ]
also first oversegmented images into superpixels. Superpixels were labeled under a
mixture of CRF. Images in a database were grouped into several contexts and each
context was modeled by a separate CRF.
Torralba et al. [ 45 ] proposed Boosted Random Fields for object detection and
segmentation. Boosting was used to learn the graph structure and local evidence
of a conditional random field. The graph structure of CRF was learned using

χ