UnAV-100
收藏资源简介:
UnAV-100是首个基于未剪辑视频的大型音视频数据集,由南方科技大学等机构创建。该数据集包含超过10,000个未剪辑视频,涵盖100种不同的音视频事件类别,总计超过30,000个事件。每个视频平均包含2.8个音视频事件,这些事件通常相互关联,可能在现实生活中同时发生。数据集广泛覆盖人类活动、音乐表演、动物/车辆/工具/自然声音等多个领域。创建过程中,从VGGSound数据集中筛选并下载原始视频,经过手动验证和标注,确保事件的准确性和相关性。UnAV-100数据集旨在推动复杂音视频视频理解的探索,特别是在音视频事件的密集定位和识别方面,解决实际应用中的挑战。
UnAV-100 is the first large-scale untrimmed audio-visual dataset based on unedited videos, created by institutions including Southern University of Science and Technology and others. This dataset contains over 10,000 untrimmed videos, covering 100 distinct audio-visual event categories, with a total of more than 30,000 events. Each video contains an average of 2.8 audio-visual events, which are usually interrelated and may occur simultaneously in real-life scenarios. The dataset extensively covers multiple domains such as human activities, musical performances, animal, vehicle, tool and natural sounds. During the creation process, raw videos were screened and downloaded from the VGGSound dataset, followed by manual verification and annotation to ensure the accuracy and relevance of the events. The UnAV-100 dataset aims to promote research on complex audio-visual video understanding, particularly in dense localization and recognition of audio-visual events, to address challenges in practical applications.




