Today, film analysis is still a tedious process performed mostly manually by film experts. Existing computer vision approaches aim at improved retrieval and summarization methods rather than at film understanding. While current research is predominantly focused on the question what can we learn and extract from a film as the final product, this thesis aims at the study of the filmmaking process as a source for high-level content information. The central question of this thesis is what can computer vision methods provide to support film analysis as performed by film expert? We discuss a possible mapping between factors that influence the production, presentation, and perception of movies, their application by means of well-established film techniques, and existing feature extraction methods in computer vision. This novel view on film analysis allows for the exploration and identification of three areas in the domain of automated film analysis and understanding. The first area comprises research tasks that have been subject to active research in the recent past. The second area covers research topics that are not immediately solvable for a fully automated computer vision approach without any prior knowledge. The last area identifies research tasks that are still open in the context of automated film analysis and understanding. Finally, we introduce three novel research questions and possible solutions: camera take reconstruction, film comparison, and recurring element detection. Performed experiments reveal two significant potentials. First, they can assist film experts by providing support for tasks that are currently performed manually. Second, proposed algorithms blaze the trail for advanced application scenarios such as the analysis of different montage patterns, the identification of missing shots, the reconstruction of the original film cut, or the detection of recurring elements.

Top