The main motivation of this work is to find and evaluate solutions for generating binary masks (silhouettes) of foreground targets in an automatic way. To this end, four renowned unsupervised image segmentation algorithms are applied to foreground segmentation. A comparison among these algorithms is carried out using the MuHAVi dataset of multi-camera human action video sequences. This dataset presents significant challenges in terms of harsh illumination resulting for example in high contrast and deep shadows. The segmentation results have been objectively evaluated against manually derived ground-truth silhouettes.
展开▼
