In automatic film and video analysis, Shot Boundary Detection (SBD) and Shot Type Classification (STC) are fundamental pre-processing steps. While previous research focuses on detecting and classifying shots in different genres such as sports movies, documentaries or news clips, only few studies investigate SBD and STC in digitized historical film footage. The main aim of this paper is to present first results on the fundamental topics SBD and STC in the context of the Horizon 2020 project Visual History of the Holocaust (VHH).
A deep learning-based SBD approach has been implemented to detect Abrupt Transitions (ATs). Furthermore, a CNN-based algorithm has been analyzed and optimized in order to classify shots into the four categories: Extreme-Long-Shot (ELS), Long-Shot (LS), Medium-Shot (MS) and Close-Up (CU). Finally, both algorithms have been evaluated on a self-generated historical dataset related to the National Socialism and the Holocaust.
The outcome of this paper demonstrates a first quantitative evaluation of the SBD approach and displays a F1;Score of 0.866 without the need of any re-training or optimization. Moreover, the proposed STC algorithm reaches an accuracy of 0.71 on classifying shots. This paper contributes a significant base for future research on automatic shot analysis related to the VHH project.