Temporal video alignment based on integrating multiple features by adaptive weighting

This paper proposes a method for establishing temporal correspondence between two different videos of the same manufacturing processes to enable visual comparison of the work in progress to support the transfer of expertise and skills. We extract two features from work objects and the hand motions of workers from the videos and align the two videos by integrating these features. For integrating the features, we propose a method that adaptively weights the two features to obtain the distance between frames. Using pairs of videos of PC assembly and cooking tasks, we show that our method offers better frame-by-frame alignment accuracy than the methods that employ each feature separately.