Detection of important segments in cooking videos

The paper presents a method to delineate visually important segments from cooking videos, which is expected to be a core technique for cooking video indexing. Visual information is exceptionally important in cooking videos, since it represents essential technical skills owing to properties of instruction type videos. We reveal that an ordinary shot-based video indexing approach will not work for cooking videos, since visually important segments are absorbed in a shot. Based on the observation of cooking videos, the method detects repetitious motion segments as visually important segments by using relatively simple and robust techniques. The experimental evaluation shows that the method detects more than 80% of manually selected segments, while 84% of detected segments were correct. We also develop a cooking video abstraction system as a sample application of our method and confirm that the proposed method is useful for a real application