inclusionAI/MemDreamer
收藏资源简介:
MemDreamer数据集包含预构建的分层图记忆和智能检索推理轨迹,用于支持长视频理解任务。该数据集与MemDreamer方法相关,方法通过解耦感知和推理,利用视频流构建分层图记忆(包括视频根、超级事件、宏观事件以及实体/事件子图),并采用基于工具增强的智能检索机制(通过观察-推理-行动循环)来回答问题。数据集可用于重现MemDreamer在多个基准测试(如LVBench、LongVideoBench等)中的最先进结果,帮助用户评估和扩展长视频理解能力。数据集包括记忆文件和推理轨迹,用户需下载后配合提供的代码进行使用。
The MemDreamer dataset includes pre-built hierarchical graph memory and agentic retrieval inference trajectories for long video understanding tasks. It is associated with the MemDreamer method, which decouples perception and reasoning by streaming videos to construct a hierarchical graph memory (comprising Video Root, Super Events, Macro Events, and Entity/Event Subgraphs) and employs tool-augmented agentic retrieval via an Observation-Reason-Action loop to answer questions. The dataset enables reproduction of MemDreamers state-of-the-art results across multiple benchmarks (e.g., LVBench, LongVideoBench), supporting evaluation and extension of long video understanding capabilities. It consists of memory files and inference trajectories, which users can download and use with the provided code.




