DAVE
收藏资源简介:
DAVE(Diagnostic Audio Visual Evaluation)是一个新颖的基准数据集,由KU Leuven的ESAT-PSI研究机构创建,旨在系统地评估音频视觉模型的性能。该数据集包含2426个样本,通过半自动数据生成方法,利用Epic Kitchens和Ego4D两个数据集生成多选问题和答案。DAVE特别设计的问题需要同时利用音频和视觉模态的信息,以确保单一模态无法正确回答问题。数据集覆盖了多种日常活动和声音事件,通过精确控制视觉动作与合成音频事件的时间对齐,以严格评估音频视觉集成能力。
DAVE (Diagnostic Audio Visual Evaluation) is a novel benchmark dataset developed by the ESAT-PSI research group at KU Leuven, designed to systematically evaluate the performance of audio-visual models. It contains 2426 samples generated via a semi-automatic data generation pipeline that leverages the Epic Kitchens and Ego4D datasets to create multiple-choice questions and their corresponding answers. The bespoke questions tailored for DAVE require joint utilization of both audio and visual modalities, ensuring that no single modality can correctly answer the questions. The dataset covers a wide range of daily activities and sound events, with precise temporal alignment between visual actions and synthesized audio events to enable rigorous assessment of audio-visual integration capabilities.




