Object-based technology is important for computer vision applications including gesture understanding, image recognition, augmented reality, etc. However, extracting the shape information of semantic objects from video sequences is a very difficult task, since this information is not explicitly provided within the video data. Therefore, an application for extracting the semantic video object is indispensable and important for many advanced applications.
An algorithm for semi-automatic video object extraction system has been developed. The performance measures of video object extraction system; including evaluation using ground truth and error metric is shown, followed by some practical uses of our video object extraction system.
The principle at the basis of semi-automatic object extraction technique is the interaction of the user during some stages of the segmentation process, whereby the semantic information is provided directly by the user. After the user provides the initial segmentation of the semantic video objects, a tracking mechanism follows its temporal transformation in the subsequent frames, thus propagating the semantic information.
Since the tracking tends to introduce boundary errors, the semantic information can be refreshed by the user at certain key frame locations in the video sequence. The tracking mechanism can also operate in forward or backward direction of the video sequence.
The performance analysis of the results is described using single and multiple key frames; Mean Error and “Last_Error”, and also forward and backward extraction. To achieve best performance, results from forward and backward extraction can be merged.
Keywords: forward and backward semi-automatic video object extraction, performance evaluation, multiple key frames.