Analysis of methods for border ownership estimation

Olgunsoylu, Sertaç
Border Ownership (BO) signifies which side of an image border owns the border. This information is very important for many computer vision problems. There are various approaches for estimating BO in the literature. One approach is to use methods that rely on low-level local features; combinations of local visual cues such as convexity or junctions and using spectral features have been proven to be quite effective. Alternatively, global constraints or cues can be imposed on top of local cue predictions to make better BO estimates. There are also physiologically-inspired attempts, which, e.g., model contrast surround modulation in complex cells to estimate BO. In this thesis, we analysed representative BO estimation methods using a comprehensive BO database, which includes 500 indoor and 500 outdoor images. Performance of each method is analysed individually in terms of estimation accuracy. In addition to this analysis, we improve a naive local cue combination method using Conditional Random Fields to impose global consistency on junctions. Moreover, we proposed a new method that uses AdaBoost algorithm to combine visual cues.
Citation Formats
S. Olgunsoylu, “Analysis of methods for border ownership estimation,” M.S. - Master of Science, 2015.