Director in a box: learning cinematic rhetoric for camera shot selection
Munro, John Burnet
MetadataShow full item record
Automatic generation of cinematic content has been a goal for both the military and the entertainment industry to allow more diverse plot structures so that a trainee or player may have a scenario tailored to their personal needs and desires. We approach this problem from a traditional view of story as being appropriately broken into two parts: plot and discourse. We focus on the rhetorical aspects of discourse, specifically selecting coherent and aesthetically pleasing shot and blocking constraints for a virtual cinematographer. In the past, selection has been solved using a decompositional planning approach. Unfortunately, each decompositional unit corresponding to a single film idiom must be hand-authored by an expert cinematographer, resulting in an intractable knowledge acquisition problem, prone to error and subjectivity. We show that this problem can instead be solved by reinforcement learning techniques, which train on features from existing sitcom and movie scenes. We will evaluate the precision of our method by running 10-fold cross validation on our training sets.