The visual movie analytics workflow combines automatic text and video preprocessing with interactive multilayer timeline visualisation to support iterative reasoning about movie content. The workflow begins with automatic preprocessing: optical-flow-based shot detection segments the video into shots; a movie script is parsed to extract characters, locations, and dialogue structure; script elements are temporally aligned with the movie via subtitle matching using cascaded similarity measures; and scenes are delineated by combining script structure with detected shot boundaries. Scene-level features — motion histograms, tf-idf text distributions, and semantic frames — are then computed to enable multiple similarity comparison methods.
The results are presented in an interactive multilayer timeline visualisation where individual elements (characters, keywords) appear as colour-coded occurrence bars across scenes (Layer 1), with expandable shot-level detail (Layer 2) and storyboard frames (Layer 3). From this overview, the analyst identifies main characters and initial story structure. The analyst then enters an iterative reasoning loop: formulating queries (keyword search, element combinations, or motion/text/semantic-frame similarity against a reference scene), inspecting results as new timelines, comparing selected scenes at finer granularity, and identifying content patterns such as character constellations, thematic events, and narrative relationships. Identified patterns are documented as new named element tags (group elements, custom annotations) that become available for subsequent queries, enabling progressive derivation of higher-level insights. The cycle continues until the analyst has developed sufficient understanding of the movie's content — who is involved, what happens, where it occurs, and when — at which point findings are synthesised into a coherent account of character dynamics, events, and thematic structure.