Extraction of read text using a wearable eye tracker for automatic video annotation

This paper presents an automatic video annotation method which utilizes the user's reading behaviour. Using a wearable eye tracker, we identify the video frames where the user reads a text document and extract the sentences that have been read by him or her. The extracted sentences are used to annotate video segments which are taken from the user's egocentric perspective. An advantage of the proposed method is that we do not require training data, which is often used by a video annotation method. We examined the accuracy of the proposed annotation method with a pilot study where the experiment participants drew an illustration reading a tutorial. The method achieved 64.5% recall and 30.8% precision.