Digital Signal Processing Reference
In-Depth Information
Fig. 5.2
Region-based and object-based multiview image segmentation
5.2.1
Region-Based Multiview Image Segmentation
The performance of understanding the semantic content of image and dynamic
scene can be improved by the region-based segmentation and classification from
multiple view images. Xiao and Quan [
44
] proposed a simple and powerful mul-
tiview segmentation framework, on the Google Maps Street View images captured
by a camera mounted on a car driving along the street. A Markov Random Field
is defined with associated graph for the multiple images in the same sequence,
where a node in the graph corresponds with a super-pixel by over-segmentation.
The extracted 2D image-based appearance and position features, as well as the 3D
geometric features are collected to learn the AdaBoost classifiers for each class la-
bel and define the unary potential function in the Gibbs energy. Color difference in
the same image and dense corresponds across different views are utilized to enforce
smoothness and consistency. The segmentation results apply to the recognition of
street view scene containing several semantic classes such as ground, building, sky,
vehicle and person.
Decomposing the image into multiclass regions is a challenging task due to
various visual concepts involved. In certain applications, the users are more in-
terested in accessing a specific object rather than the scene, which makes the
