Hide/Show Apps

Estimating border ownership using iterative vector voting and conditional random fields

Özkan, Buğra
Border ownership is the information that signifies which side of a border owns the border. Estimating this information has recently become very popular for perceptual organization as it allows rectification of ambigious visual information. It is applied on many computer vision problems such as object detection, depth perception and optical flow. In this thesis, two different approaches are followed to solve the border ownership problem. For the supervised approach, conditional random fields are used as it is the most appropriate method for modelling con- textual relations between semantic classes. Tensor voting is the inspire of our second algorithm called Iterative Vector Voting, as it allows modelling different information sources and their interactions. It is an unsupervised voting frame- work, which is proper for the use of Gestalt visual cues. Experiments show that both two models show significant contribution to the border ownership problem with respect to the successful results gathered on our own large-scale dataset.