MER2024

该数据集名为MER2024，专为情感识别挑战而设计，包含四个子集：训练与验证集、半监督集、噪声集和重叠语音集。其中，训练与验证子集提供了5030个带标签的单说话人视频片段，用于训练和验证；半监督子集包含了115,595个未带标签的数据点，以及1169个作为真实测试集的数据点，以评估性能表现。该数据集涵盖了六种离散的情感标签：担忧、快乐、中性、愤怒、惊讶和悲伤。挑战鼓励使用半监督学习技术。该数据集规模庞大，任务为情感识别。

This dataset, designated as MER2024, is specifically developed for emotion recognition challenge tasks. It encompasses four subsets: training and validation subset, semi-supervised subset, noisy speech subset, and overlapping speech subset. The training and validation subset provides 5030 labeled single-speaker video clips for model training and validation. The semi-supervised subset contains 115,595 unlabeled data samples, alongside 1169 data points that serve as the ground-truth test set for performance evaluation. This dataset covers six discrete emotion labels: anxiety, happiness, neutral, anger, surprise, and sadness. The corresponding challenge encourages the application of semi-supervised learning techniques. As a large-scale dataset, it centers on the task of emotion recognition.

搜集汇总

背景与挑战

背景概述

MER2024-SEMI是一个专注于多模态情感识别的数据集，采用对比学习和自训练方法解决有限标注样本问题，在官方评测中取得了88.25%的WAF成绩。数据集由内蒙古大学、陕西师范大学和南京大学联合开发，主要用于情感计算研究领域。