RareAct

RareAct是由ENS/Inria创建的一个视频数据集，专注于收集不寻常的人与物体交互动作，如‘blend phone’, ‘cut keyboard’等。该数据集包含122种不同的动作，通过结合罕见的动词和名词组合而成，这些组合在大型文本语料库HowTo100M中很少共同出现。数据集的创建过程涉及从视频分享平台搜索相关视频，并将其分割成10秒的片段进行人工标注。RareAct主要用于评估和提升动作识别模型在处理罕见动作组合时的组合能力，特别是在零样本和少量样本学习场景下的表现。

RareAct is a video dataset developed by ENS/Inria, which focuses on curating unusual human-object interaction actions such as 'blend phone', 'cut keyboard' and other similar cases. This dataset includes 122 distinct action categories, which are formed by combining rare verb-noun pairs that rarely co-occur in the large-scale text corpus HowTo100M. The dataset construction process involves searching for relevant videos from video-sharing platforms, segmenting them into 10-second clips and conducting manual annotation. RareAct is primarily designed to evaluate and enhance the compositional generalization capability of action recognition models when handling rare action combinations, especially their performance in zero-shot and few-shot learning scenarios.

搜集汇总

背景与挑战

背景概述

RareAct是一个由ENS/Inria创建的视频数据集，专注于收集122种不寻常的人与物体交互动作，如'blend phone'和'cut keyboard'，这些动作基于罕见的动词和名词组合，源自HowTo100M语料库中很少共同出现的组合。数据集通过从视频分享平台搜索并分割成10秒片段进行人工标注，主要用于评估和提升动作识别模型在处理罕见动作组合时的组合能力，特别是在零样本和少量样本学习场景下。