A Bayesian approach to video object segmentation via merging 3-D watershed volumes