Caltech Mouse Social Interactions 2021 (CalMS21) Dataset

CalMS21数据集是由加州理工学院创建的行为神经科学领域的多代理行为数据集。该数据集包含轨迹数据，记录了自由行为的小鼠在标准居民-入侵者测定中的社会互动视频。数据集提供了三种设置下的基准，以评估自动化行为分类方法的性能：(1) 单注释者标注的大规模行为数据集的训练，(2) 学习行为定义中的注释者间差异的风格转移，(3) 在有限训练数据下学习新感兴趣的行为。数据集包含600万帧未标记的跟踪姿势数据，以及超过100万帧带有跟踪姿势和相应帧级行为注释的数据。该数据集旨在解决能够准确使用标记和未标记跟踪数据进行行为分类的挑战，以及能够推广到新设置的能力。

The CalMS21 dataset is a multi-agent behavioral dataset developed by the California Institute of Technology for behavioral neuroscience research. This dataset includes trajectory data and videos documenting social interactions of freely behaving mice in the standard resident-intruder assay. The dataset provides benchmarks under three experimental setups to evaluate the performance of automated behavioral classification methods: (1) training on large-scale behavioral datasets annotated by a single annotator; (2) style transfer to learn inter-annotator discrepancies in behavioral definitions; (3) learning novel behaviors of interest under limited training data. The dataset contains 6 million frames of unlabeled tracked pose data, as well as over 1 million frames of data with tracked poses and corresponding frame-level behavioral annotations. This dataset aims to address the challenges of accurately utilizing both labeled and unlabeled tracked data for behavioral classification, as well as the capability to generalize to novel settings.