Unsupervised Graph Based Video Object Extraction

A method to extract the object containing regions in the ‘object proposal’ set in the video is employed. The segmented object containing areas are then employed to construct segmentation models for optimal video object extraction. First, an unsupervised graph based framework is used for detection and extraction of foreground in the video. We take into account the general properties (spatially cohesiveness, smooth motion trajectories, predicted-shape similarity, appearance and motion across frames) of object and use this to extract object from all available object proposals. Second, object proposal expansion is done by using motion based proposal predictions and unsupervised graph is constructed based on these proposals. Last, a discriminative function is introduced to differentiate between moving and background. This method is evaluated on segtrack dataset and it gives better result than the state-of-the-art.