ResearchPod Summary
This paper addresses the challenge of creating engaging, data-driven narratives for sports videos. While prior work has focused on embedding static visualizations or simple overlays into sports footage, there is a lack of tools that allow creators to build cohesive, time-synced stories that integrate data, camera transitions, and race events. The authors focus on swimming as a case study due to its rich, dynamic data and global popularity.
To solve this, the researchers developed an automated pipeline that ingests raw race videos, commentary audio, and event metadata. Using computer vision (SAM3) for lane-based tracking and Large Language Models (LLMs) for processing commentary, the pipeline generates structured, time-aligned data. This data powers SwimComposer, a technology probe that allows creators to author narrative visualizations—such as athlete introductions, leader tracking, and chase comparisons—that move dynamically with the athletes on screen.
The authors evaluated SwimComposer with experienced content creators and graphic designers. The study found that the tool successfully enables the creation of narrative-driven race videos by providing a structured way to coordinate data layers, camera shifts, and pacing. Participants were able to move beyond simple data overlays to create more complex, story-driven content. The study also identified design challenges, such as the need for better automated narrative suggestions and the difficulty of balancing data density with visual clarity in fast-paced sports.
As sports content becomes increasingly consumed on social media, there is a growing demand for videos that do more than just show the raw event. By automating the labor-intensive data preparation process and providing a framework for narrative authoring, this work lowers the barrier for creators to produce high-quality, informative sports content. It shifts the focus from merely embedding charts to crafting a deliberate visual story that helps audiences understand the nuances of a competition.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.