FERV39k

我们建立了一个大规模的多场景数据集，称为FERV39k。我们从三个方面分析了构建此类新颖数据集的重要组成部分 :( 1) 多场景层次结构和表达式类; (2) 候选视频剪辑的生成; (3) 可信的手动标记过程。基于这些准则，我们选择了4个场景，分为22个场景，根据精心设计的工作流，对4k视频自动获得的86k样本进行注释，最终构建38,935个带有7个经典表情的视频片段。还提供了四种基准框架的实验基准，并进一步分析了它们在不同场景中的性能，并给出了未来研究的一些挑战。

We introduce a large-scale multi-scenario dataset named FERV39k. We analyze the core components for constructing such a novel dataset from three aspects: (1) multi-scenario hierarchy and expression categories; (2) generation of candidate video clips; (3) reliable manual annotation process. Based on these guidelines, we selected 4 scenarios which were further divided into 22 sub-scenarios. Using a well-designed workflow, we annotated 86k automatically acquired samples from 4K videos, and finally constructed 38,935 video clips containing 7 classic facial expressions. We also provide experimental benchmarks for four baseline frameworks, further analyze their performance across different scenarios, and present several challenges for future research.

搜集汇总

背景与挑战

背景概述

FERV39k是一个大规模多场景面部表情识别数据集，包含38,935个视频片段，涵盖7种经典表情，从4k视频中自动注释生成。该数据集由复旦大学于2022年发布，提供了多场景层次结构和基准框架实验，适用于面部表情识别研究。