SF20K competition shows narrative understanding, not model size, is key for video QA

The first Short-Films 20K (SF20K) Competition, held alongside ICCV 2025, focused on advancing story-level video understanding through an open-ended question-answering task. Using a benchmark of amateur short films and evaluated by GPT-4.1-nano, the competition saw 22 teams submit entries. Analysis of the results indicates that narrative-aware, shot-level processing and multi-stage pipelines are more effective than simple frame sampling, and that subtitle quality significantly impacts performance. AI

IMPACT Highlights that information selection and reasoning structure, rather than raw model capacity, are key challenges in long-form video question-answering.

RANK_REASON This is a summary of findings from a competition detailed in an academic paper. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• GPT-4.1-nano
• ICCV 2025
• SF20K Competition
• Short-Films 20K

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →