Dynamic Hand Gesture Recognition from Egocentric Videos based on SlowFast Architecture

Recently, thanks to a large number of lightweight digital recording devices used in different applications, the amount of egocentric data has increased overtime. Compared with videos captured by ambient cameras, egocentric videos have their own challenges as they may contain large, non-linear and unpredictable motion. This paper presents an approach for hand gesture recognition from egocentric videos based on SlowFast network architecture. The model involves Slow and Fast pathway in which a Slow pathway operates at low frame rate while a Fast pathway performs at high frame rate. In egocentric videos, some hand movements happen faster or slower than others depending on the actor or surrounding context. In order to retain egocentric-based attribute of the video, we perform extensive experiments on two branches of information by dividing input frames into Slow pathway and Fast pathway. As a result, our method has achieved better classification accuracy scores in EgoGesture dataset compared with other state-of-the-art frameworks such as VGG-16+LSTM, C3D+LSTM+RSTTM models.