TemporalBench
收藏资源简介:
TemporalBench是一个专注于评估多模态视频模型细粒度时间理解能力的新基准数据集。该数据集由约10,000个视频问答对组成,这些问答对来源于约2,000个高质量的人工标注,详细描述了视频片段中的时间动态。数据集的内容涵盖了从细粒度动作理解到长视频理解的多种任务,旨在解决现有视频基准在时间理解评估上的不足。数据集的创建过程包括从现有视频基准中收集视频片段,并通过亚马逊机械土耳其工人进行详细标注,然后利用大型语言模型生成负样本并进行筛选。TemporalBench的应用领域广泛,包括视频问答、视频字幕生成、长视频理解等,旨在推动模型在时间推理能力上的研究。
TemporalBench is a novel benchmark dataset dedicated to evaluating the fine-grained temporal understanding capabilities of multimodal video models. It comprises approximately 10,000 video question-answer pairs derived from around 2,000 high-quality manual annotations that meticulously describe the temporal dynamics within video clips. The dataset covers a wide range of tasks spanning from fine-grained action understanding to long-form video understanding, aiming to address the shortcomings of existing video benchmarks in temporal understanding evaluation. The dataset creation process involves collecting video clips from existing video benchmarks, conducting detailed annotations via Amazon Mechanical Turk workers, then generating and screening negative samples using large language models. TemporalBench has broad application scenarios including video question answering, video captioning, long-form video understanding and others, with the goal of advancing research on the temporal reasoning capabilities of models.

- 1TemporalBench: Benchmarking Fine-grained Temporal Understanding for Multimodal Video Models微软研究院 · 2024年



