Audiovisual Moments in Time (AVMIT)

AVMIT是一个大规模视听动作事件数据集，包含57,177个视频的视听标注，由伯明翰大学计算神经科学与认知机器人中心创建。数据集中的每个视频都经过至少3名训练有素的参与者独立评估，确保标注质量。AVMIT特别适用于研究视听对应关系，支持深度神经网络和人类参与者的研究和比较实验。此外，数据集还提供了两套预计算的视听特征嵌入，降低了视听深度神经网络研究的门槛。

AVMIT is a large-scale audio-visual action event dataset containing audio-visual annotations for 57,177 videos, created by the Centre for Computational Neuroscience and Cognitive Robotics at the University of Birmingham. Every video in the dataset has been independently evaluated by at least three trained participants to ensure annotation quality. AVMIT is particularly well-suited for researching audio-visual correspondence, and supports research and comparative experiments involving both deep neural networks and human participants. Additionally, the dataset provides two sets of pre-computed audio-visual feature embeddings, lowering the barrier to entry for research on audio-visual deep neural networks.