Multi-Event Causal Discovery (MECD)
收藏资源简介:
Multi-Event Causal Discovery (MECD) 数据集由上海交通大学的研究团队创建,旨在解决视频因果推理中的多事件因果发现任务。该数据集包含多个按时间顺序排列的事件片段及其对应的文本描述,要求识别这些事件之间的因果关系,并生成一个结构化的因果图。数据集来源于多个广泛使用的长期日常视频数据集,如ActivityNet Captions、EgoSchema和NExTVideo。数据集的构建过程包括手动标注事件对之间的因果关系,以支持模型训练和评估。该数据集的应用领域包括视频问答和视频事件预测,旨在通过因果推理提升视频理解能力。
Multi-Event Causal Discovery (MECD) dataset was developed by a research team from Shanghai Jiao Tong University, targeting the multi-event causal discovery task in video causal reasoning. This dataset contains multiple temporally ordered event clips and their corresponding text descriptions, requiring models to identify the causal relationships between these events and generate a structured causal graph. The dataset is sourced from several widely used long-form daily video datasets, such as ActivityNet Captions, EgoSchema and NExTVideo. Its construction process includes manual annotation of causal relationships between event pairs to support model training and evaluation. The application fields of this dataset cover video question answering and video event prediction, aiming to improve video understanding capabilities through causal reasoning.

- 1MECD+: Unlocking Event-Level Causal Graph Discovery for Video Reasoning上海交通大学 · 2025年



