Effective Comic-Like Representations with Embedded Regions of Interest

Comic-like summaries exploit the narrative structure of comics to create intuitive and easily readable abstracts. However, real comics use complex composition techniques which are difficult to mimic in an unsupervised way as they involve high level semantic understanding. This paper explores the use of visual attention analysis and face detection to embed regions of interest in adjacent images, obtaining more compact yet informative representations. This paper also addresses the generation of the layout, which involves combinatorial optimization problems. In practice, using exhaustive search to solve the problem is not feasible due to the large number of images. A split and merge approach is proposed to effectively address the layout problem, thus the limitations of finding solutions in a wide range of row widths can be avoided. A user study conducted on several episodes of TV series confirmed the utility of the proposed approach.