Video sharing over the Internet (on sites like YouTube and Vimeo) is on the rise, and many members of the sign language community use these sites to share sign language videos. There has been a rapid increase in the video data available on these sites; this appeals to computer scientists developing effective video search tools for easy and quick access.

The authors propose a support vector machine (SVM)-based classifier method to search sign language videos, extracting video features using background modeling and face detection. The background model compares each gray-scale pixel value in consecutive video frames and assigns the pixel in the later frame as a foreground pixel if their difference is more than a threshold (the authors used 45).

The model extracts five features after face detection, which are based on the quantity of movement (two features), continuity of motion (one feature), and location of motion (two features). The authors claim that combining the object’s motion symmetry and position of the object’s face in a video is a measure giving the best performance of any single feature.

For classifier evaluation, the paper uses 192 videos, including 98 sign language videos (78 in American Sign Language and 20 in British Sign Language) and 94 non-sign language videos, from YouTube and Vimeo. The authors randomly selected training and test data to test the classifier performance based on 1,000 executions measuring precision, recall, and F1 score.

The authors propose to increase the accuracy of the final foreground objects for sign language video and interface detection in subsequent frames as future enhancements. This paper is an interesting read on sign language video classification and search on video-sharing sites.