文章背景与核心概要
基于图的检索增强生成（RAG）是长视频理解的强大工具，但它往往受到固定时间粒度的困扰——迫使系统在索引和推理阶段使用相同级别的细节。本文介绍了一种名为“密度感知图构建”（Density-Aware Graph Construction, DAGC）的无需训练方法，成功解耦了这两个过程。
通过将视觉上冗余的视频片段合并为紧凑的粗粒度索引，DAGC在保持高效检索的同时，依然能够在需要时扩展回细粒度证据以进行精准推理。该方法显著提升了性能效率，实现了高达 \(1.7\times\) 的加速，同时保留了99%的原始问答准确率。
Coarse Indexing, Fine Evidence: Decoupling Temporal Granularity in Long-Video RAG
arXiv: 2608.23011
Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
Authors: Zhe Jin, Zhimin Lin, Bin Zheng, Junhua Fang, Huihua Yang
Submitted: 24 Aug 2026
Coarse Indexing, Fine Evidence: Decoupling Temporal Granularity in Long-Video RAG
arXiv: 2608.23011
Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
Authors: Zhe Jin, Zhimin Lin, Bin Zheng, Junhua Fang, Huihua Yang
Submitted: 24 Aug 2026
摘要
基于图的检索增强生成（RAG）是长视频理解的强大工具，但它往往受到僵化的时间粒度限制——迫使系统在索引和推理阶段采用相同的细节层级。
作者引入了密度感知图构建（Density-Aware Graph Construction, DAGC），这是一种无需训练的方法，能够将这两个过程进行解耦。通过将视觉冗余的视频片段合并为紧凑的粗粒度索引，DAGC在保持高效检索的同时，保留了扩展回细粒度证据以进行精确推理的能力。该方法显著提升了性能效率，在保留99%原始问答准确率的同时，实现了高达 \(1.7\times\) 的加速。
Summary
Graph-based Retrieval-Augmented Generation (RAG) is a powerful tool for long-video understanding, but it often suffers from rigid temporal granularity—forcing the system to use the same level of detail for both indexing and reasoning.
The authors introduce Density-Aware Graph Construction (DAGC), a training-free approach that decouples these two processes. By merging visually redundant video segments into a compact, coarse index, DAGC allows for efficient retrieval while maintaining the ability to expand back to fine-grained evidence for precise reasoning. This method significantly improves performance efficiency, achieving up to \(1.7\times\) acceleration while retaining 99% of original QA accuracy.
核心贡献
- 粒度解耦： 挑战了在索引和证据检索中保持相同时间分辨率的必要性。
- 密度感知图构建（DAGC）： 一种新颖的、无需训练的机制，通过识别并合并冗余的时间区块来创建紧凑的图索引。
- 效率提升： 将图节点数量减少了 50–60%，并实现了显着的端到端实际运行时间加速。
- 通用性： 证明了在各种大视觉语言模型（LVLM）骨干网络和现有RAG管道中均能保持一致的性能提升。
Key Contributions
- Decoupling Granularity: Challenges the necessity of maintaining identical temporal resolution for indexing and evidence retrieval.
- Density-Aware Graph Construction (DAGC): A novel, training-free mechanism that creates a compact graph index by identifying and merging redundant temporal chunks.
- Efficiency Gains: Reduces the graph node count by 50–60% and provides substantial end-to-end wall-clock acceleration.
- Versatility: Demonstrates consistent performance improvements across various Large Vision-Language Model (LVLM) backbones and existing RAG pipelines.
实验结果
DAGC的有效性在三大主流基准测试中得到了验证： * MLVU * VideoMME * LongVideoBench
性能亮点： * 图优化： 仅保留了原始图节点的 40–50%。 * 速度： 实现了 \(1.3\times\) 到 \(1.7\times\) 的端到端实际运行时间加速。 * 准确率： 保留了大约 99% 的原始问答性能，证明了粗粒度索引并不会以牺牲推理质量为代价。
Experimental Results
The effectiveness of DAGC was validated across three major benchmarks: * MLVU * VideoMME * LongVideoBench
Performance Highlights: * Graph Optimization: Retains only 40–50% of original graph nodes. * Speed: Achieves \(1.3\times\) to \(1.7\times\) end-to-end wall-clock acceleration. * Accuracy: Preserves approximately 99% of the original QA performance, proving that coarse indexing does not compromise reasoning quality.
访问与资源
Access & Resources