UDMT dataset: Manual annotations for behavioral recordings used in unsupervised multi-animal tracking

This archive contains the manually annotated ground-truth trajectories for the UDMT behavioral recording datasets. The corresponding behavioral videos are provided at Zenodo record 14580256: https://zenodo.org/records/14580256. The annotation files are stored in MOT format and can be matched to the videos according to their dataset/video names. Each annotation file follows the standard MOT-style frame-level format, where each row corresponds to one annotated animal instance in one annotated frame. The annotations provide object identities and bounding-box locations for evaluating multi-animal tracking performance. The manual annotations were generated at fixed frame intervals rather than for every video frame. The annotation interval depends on the dataset name: - Datasets containing "94hz": one annotated frame every 600 frames. - Datasets containing "flies": one annotated frame every 150 frames. - Datasets containing "celegans": one annotated frame every 150 frames. - All other datasets: one annotated frame every 300 frames. The dataset names and video names are designed to correspond directly by filename. These annotations are intended for quantitative evaluation of tracking accuracy and identity consistency in UDMT and related multi-animal tracking methods.