DFEW真实场景动态表情数据库

DFEW（Dynamic Facial Expression in the Wild）是本项目建立的大规模真实场景动态表情数据库，采集了16372个非受限场景中包含七种基本情绪（开心、伤心、中性、愤怒、惊讶、厌恶、害怕）的动态表情视频片段，每个样本都由10名不同的受过表情标注训练的标注者标注了共计10次。具体建库流程如下：首先，我们运用网络爬虫技术，从互联网上收集了超过1500部高分辨率的、与生活密切相关的电影，将其作为获取面部动态表情视频片段的来源。电影包括各种题材：喜剧、悲剧、战争、爱情等等。然后，我们雇佣了数十名学生，使用视频剪辑软件，从电影中人工剪辑包含七种基本表情的人脸视频片段，共计16372段。其次，我们雇佣了众包公司的12名标注人员，并在正式标注前对其进行了专业的表情认知的训练。在此基础上，要求他们对上述采集的人脸视频片段重新进行七种基本表情最相近的情感标签，每个视频都由其中的10人独立标注10次，最终获得16372个视频片段样本的七维情感向量标签。我们将最高投票数过半（大于等于6票）的表情样本组成DFEW数据库单标签子集（赋予七种基本情绪之一的标签），单标签子集共计12059个视频样本；剩余4313个样本组成DFEW数据库的多维标签子集。DFEW数据库可以提供单标签动态表情识别、多维标签表情识别等任务的基准测试。

DFEW (Dynamic Facial Expression in the Wild) is a large-scale real-world dynamic facial expression database constructed in this project. It contains 16,372 dynamic facial expression video clips covering seven basic emotions (happiness, sadness, neutral, anger, surprise, disgust, fear) captured in unconstrained real-world scenarios. Each sample was annotated 10 times in total by 10 distinct annotators who had received professional facial expression annotation training.

The specific database construction process is as follows: First, we used web crawler technology to collect over 1,500 high-resolution, life-related movies from the Internet as the source of facial dynamic expression video clips. These movies cover various genres, including comedy, tragedy, war, romance and more. Then, we hired dozens of students to manually clip human facial video segments containing the seven basic emotions from these movies using video editing software, resulting in a total of 16,372 clips. Next, we hired 12 annotators from a crowdsourcing company and provided them with professional facial emotion cognition training before formal annotation. On this basis, they were required to assign the emotional label that most closely matches one of the seven basic emotions to each of the previously collected facial video clips. Each video was independently annotated 10 times by 10 of these annotators, ultimately obtaining seven-dimensional emotional vector labels for all 16,372 video clip samples.

We formed the single-label subset of the DFEW database by selecting samples with no less than 6 votes (i.e., more than half of the 10 annotations) as their dominant emotion, assigning them a label corresponding to one of the seven basic emotions; this single-label subset contains a total of 12,059 video samples. The remaining 4,313 samples constitute the multi-label subset of the DFEW database.

The DFEW database can provide benchmark tests for tasks such as single-label dynamic facial expression recognition and multi-label facial expression recognition.

搜集汇总

背景与挑战

背景概述

DFEW真实场景动态表情数据库是一个大规模、高质量的表情识别数据集，包含16372个从电影中剪辑的动态表情视频片段，覆盖七种基本情绪（开心、伤心、中性、愤怒、惊讶、厌恶、害怕）。每个样本由10名训练有素的标注者独立标注，确保了标签的可靠性，并划分为单标签和多维标签子集，适用于单标签和多维标签表情识别任务的基准测试。数据来源于真实场景，具有多样性和实用性，总数据量为32.38GB。