Scholarlink

<p>Video Summarization (VS) plays a key role in efficient video browsing and retrieval. With the rise of large models, a popular VS paradigm translates video frames into textual descriptions, summarizes the text, and retrieves relevant frames from that summary. Yet this “Video-Text-Video” pipeline suffers from two core issues: (1) Inconsistent Identity: frame-level descriptions, generated independently, rarely preserve consistent entity identities across frames. (2) Summarization Instability: text-based queries for frame selection are often noisy, leading to unreliable retrieval. To address these problems, we build a summarization pipeline around Global Entity Unification (GEU) and Robust Importance Scoring (RIS). A VLM produces per-frame captions, but unlike earlier efforts, we trace each object across frames and attach consistent identifiers to it. Fragmented, isolated descriptions thus become a single, object-aware text corpus that unifies the storyline. The VLM next applies a bi-level strategy to produce a summary that balances high-level themes with detailed evidence, after which a stable, multi-granular scoring mechanism pinpoints and retrieves the key video segments. Benchmarks show that the resulting summaries are both more coherent and more faithful to the source video.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Large model-assisted video summarization via global entity unification and robust importance scoring

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Donglei Chen,
  • Shaoyu Huang,
  • Xuemiao Xu,
  • Yongwei Nie,
  • Ping Li,
  • C. L. Philip Chen

摘要

Video Summarization (VS) plays a key role in efficient video browsing and retrieval. With the rise of large models, a popular VS paradigm translates video frames into textual descriptions, summarizes the text, and retrieves relevant frames from that summary. Yet this “Video-Text-Video” pipeline suffers from two core issues: (1) Inconsistent Identity: frame-level descriptions, generated independently, rarely preserve consistent entity identities across frames. (2) Summarization Instability: text-based queries for frame selection are often noisy, leading to unreliable retrieval. To address these problems, we build a summarization pipeline around Global Entity Unification (GEU) and Robust Importance Scoring (RIS). A VLM produces per-frame captions, but unlike earlier efforts, we trace each object across frames and attach consistent identifiers to it. Fragmented, isolated descriptions thus become a single, object-aware text corpus that unifies the storyline. The VLM next applies a bi-level strategy to produce a summary that balances high-level themes with detailed evidence, after which a stable, multi-granular scoring mechanism pinpoints and retrieves the key video segments. Benchmarks show that the resulting summaries are both more coherent and more faithful to the source video.