EPIC-SOUNDS
收藏资源简介:
EPIC-SOUNDS是一个大规模的音频标注数据集,源自EPIC-KITCHENS-100的音频流,专注于捕捉日常生活中的声音事件。该数据集包含78,400个分类的音频事件段,分布在44个类别中,以及39,200个未分类的音频事件段,总计117,600个音频事件段,覆盖100小时的视频材料。数据集的创建过程涉及通过众包方式对音频事件进行时间和语义标注,特别关注音频与视觉事件的时间和语义一致性差异。EPIC-SOUNDS适用于音频识别和声音事件检测等领域的研究,旨在解决现有模型在仅依赖音频信息识别动作时的局限性。
EPIC-SOUNDS is a large-scale audio annotation dataset derived from the audio streams of EPIC-KITCHENS-100, focusing on capturing sound events in daily life. This dataset contains 78,400 categorized audio event segments distributed across 44 categories, as well as 39,200 uncategorized audio event segments, with a total of 117,600 audio event segments covering 100 hours of video material. The creation of this dataset involves conducting temporal and semantic annotation for audio events via crowdsourcing, with particular attention paid to the discrepancies in temporal and semantic consistency between audio and visual events. EPIC-SOUNDS is applicable to research in fields such as audio recognition and sound event detection, aiming to address the limitations of existing models when recognizing actions relying solely on audio information.

- 1Epic-Sounds: A Large-scale Dataset of Actions That Sound牛津大学工程科学系视觉几何组 · 2023年



