MSR-VTT Adverbs

我们评估了HowTo100M副词的方法，该方法从HowTo100M中的83个任务中挖掘了副词。由于注释是从教学视频的自动转录叙述中获得的，因此它们是嘈杂的; 44% 带注释的动作副词对在视频剪辑中不可见。该数据集包含5,824个片段，其中包含72个动词和6个副词中的动作副词对注释。此数据集的明显限制是它包含的副词数量很少，因此我们从现有的视频检索数据集中创建了三个新的副词数据集: VATEX副词，msr-vtt副词和ActivityNet副词。这些包含更少的噪音和更多种类的副词。

We evaluated an adverb-mining approach for the HowTo100M dataset, which extracted adverbs from 83 tasks in HowTo100M. As the annotations were obtained from automatically transcribed narratives of instructional videos, they are noisy; 44% of the annotated action-adverb pairs are not visible in the corresponding video clips. This dataset contains 5,824 segments annotated with action-adverb pairs spanning 72 verbs and 6 adverbs. The obvious limitation of this dataset is the small number of adverbs it covers, so we created three new adverb datasets from existing video retrieval datasets: VATEX-Adverb, MSR-VTT-Adverb, and ActivityNet-Adverb. These datasets contain less noise and a wider variety of adverbs.

搜集汇总

背景与挑战

背景概述

MSR-VTT Adverbs是一个基于MSR-VTT视频检索数据集构建的副词注释数据集，旨在减少噪音并提升副词多样性，以支持细粒度动作理解。它由阿姆斯特丹大学于2022年发布，作为现有视频检索数据集的补充扩展。