Our method uses machine learning techniques together with a large-scale training dataset of manually created GIFs. The dataset consists of about 100k GIFs that people created from videos. Using this data we train a Deep Neural Network algorithm that learns to understand what makes GIFs awesome. Finally, we use this model to automatically rank video segments and generate GIFs from the best and most interesting ones.
For more technical information on how it works see our research paper at the 2016 Computer Vision and Pattern Recognition conference: Video2GIF: Automatic Generation of Animated GIFs from Video