Abstract
Binge-watching television serials often leaves viewers struggling to recall intricate plotlines by the time a new season premieres. This paper presents an automated system for generating character-focused video summaries. By integrating dynamic social network analysis to track narrative shifts with cinematic stylistic features such as shot size and musicality, the system produces highly compressed, engaging recaps. Evaluation on Game of Thrones demonstrates that these summaries effectively mitigate the "cold-start" phenomenon, helping viewers re-engage with complex stories.
Introduction
Modern television serials are characterized by continuous, complex narratives that span dozens of hours across multiple seasons. While this format encourages binge-watching, it also creates a significant cognitive hurdle: the "cold-start" phenomenon. Viewers often experience emotional and cognitive disengagement when returning to a series after a long hiatus.
Existing video summarization techniques are largely designed for standalone movies or single episodes, failing to capture the long-term narrative arcs essential for serial television. This research addresses the gap by developing an algorithm that understands both the narrative content (character interactions) and the cinematic form (filming techniques) to automatically generate personalized character catch-up videos.
Methodology
The proposed pipeline automates the extraction, scoring, and assembly of video clips into a coherent summary.
Logical Story Units (LSUs)
To ensure visual continuity, the system identifies Logical Story Units (LSUs)—self-contained micro-scenes rather than individual shots. The extraction process utilizes a matrix-based algorithm to detect recurring camera angles, effectively identifying the "180-degree rule" commonly used in dialogue scenes to maintain spatial consistency.
Dynamic Social Network Analysis
The system models the plot by tracking changes in a character's social network. By calculating the mathematical distance between "relationship vectors" across different scenes, the algorithm identifies major narrative shifts. This allows the system to partition a character's journey into distinct eras (e.g., Arya Stark’s transition from King’s Landing to Harrenhal) without requiring complex natural language processing of dialogue.
Stylistic Saliency and Selection
LSUs are further evaluated based on cinematic cues:
- Shot Size: The system uses face detection to prioritize close-ups, which are more engaging in a summary format.
- Musicality: Audio chroma vectors are analyzed to distinguish narrative-heavy background music from standard speech.
The final selection is modeled as a quadratic knapsack problem, using a greedy heuristic to maximize both the relevance of the clips to the character and the diversity of the narrative periods covered, all within a strict time constraint.
Dataset and Evaluation
The system was evaluated using the first five seasons of the HBO series Game of Thrones. This dataset includes 50 episodes of high narrative complexity.
To validate the summarization logic independently of potential errors in automated audio-visual processing, researchers used manual annotations for speaker identities and scene boundaries. All subsequent steps, including LSU extraction and summary generation, were fully automated.
Results and Discussion
The system successfully condensed 50 hours of content into character-specific summaries of 90 to 150 seconds. A user study involving 187 participants compared three summary types: "Full" (Social + Style), "Style-only", and a semi-random baseline.
- Recap Effectiveness: The "Full" summary was significantly preferred for its ability to help viewers remember the story, with over 70% of users choosing it for major characters like Arya Stark.
- Trailer Appeal: Interestingly, "Style-only" summaries—which focused on music and close-ups—were preferred as "trailers" to build excitement for upcoming episodes.
- Efficiency: Once features are pre-computed, the system generates a custom summary in under four seconds on standard hardware.
Conclusion
This research demonstrates that social network dynamics provide a robust proxy for complex plot structures. By codifying cinematic grammar into computable metrics, the system produces summaries that are both narratively coherent and visually professional. Future work will focus on full automation of speaker diarization and exploring the psychological impact of these summaries on viewer retention and emotional empathy.
Summary updated