Pig Skeleton- and Image-Based Behavior Dataset for Spatio-Temporal Behavior Classification

This dataset contains annotation files, pose-estimation data, skeleton-based temporal windows, RGB image windows, and experimental outputs used for pig behavior classification based on skeleton-based and image-based temporal representations. The dataset was created from video recordings of pigs housed in group pens under real production conditions. The released package includes CVAT annotation files, CSV annotation exports, YOLOv8n-Pose images and labels, temporal fold definitions, out-of-fold YOLOv8n-Pose prediction outputs, skeleton windows for DeGCN and KP-TCN experiments, RGB image clips/windows for SlowFast and ViTAM-SlowFast baselines, bootstrap exports, metadata files, a class mapping file, a keypoint schema, and file manifests. The dataset supports both six-class and four-class behavioral classification scenarios. The six-class scenario includes walking, running, eating, drinking, lying awake, and sleeping. In the four-class scenario, walking and running are grouped into locomotion, while lying awake and sleeping are grouped into lying. The dataset contains annotations derived from 5,300 video frames and 63,600 annotated animal instances, with 9 anatomical keypoints per animal. It also includes YOLOv8n-Pose training files, temporal 5-fold definitions, skeleton windows for T=5 and T=15 settings, and RGB image windows used by the visual baselines. The files are divided into multiple ZIP packages to facilitate upload, download, and reuse. A SHA256 checksum file is provided to verify file integrity after download.