BROWSE
Detail
Conference Paper
Query-Driven Scene Selection for Personalized Video Summarization Using Multimodal Semantic Descriptions
Cited - time in 
Share 
Share
- Authors
- Hyun-Jeong Yim, Sungjun Ahn, Jung Sun Um, Jae Hyun Seo
- Issue Date
- 2026-07
- Citation
- International Symposium on Broadband Multimedia Systems and Broadcasting (BMSB) 2026, pp.1-4
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- This paper presents a query-driven video summarization framework based on scene-level semantic decision making using a language model. Unlike conventional video summarization approaches that rely on visual saliency, keyword matching, or learned importance scores, the proposed method interprets a user query as a high-level semantic constraint that governs scene selection. Video content is first segmented into scenes, and each scene is represented by a unified textual description constructed from visual captions and aligned speech transcripts. A language model is then employed as a semantic decision engine to determine whether each scene satisfies the semantic constraints implied by the user query. Based on the resulting binary relevance decisions, query- relevant scenes are selected and concatenated in temporal order to generate a personalized summary without modifying scene boundaries or synthesizing new content. Experimental results on real-world news and movie videos demonstrate that the proposed approach achieves high recall of query-relevant scenes while effectively excluding unrelated content, particularly for abstract and context-dependent queries. These results indicate that language-model-based semantic reasoning provides an effective and flexible mechanism for personalized video summarization.
- Keyword
- Query-driven video summarization, semantic reasoning, scene selection, multi-modal video understanding, personalized media services
- KSP Keywords
- Binary relevance, Decision-making, Flexible mechanism, High recall, Keyword matching, Language Models, Multi-modal, Personalized video, Query-driven, Real-world, Semantic constraints