SwinePose: A Lateral-View Benchmark Dataset for Pig Pose Estimation and Gait Analysis

SwinePose is the first publicly available lateral-view benchmark dataset for markerless pig pose estimation and gait analysis. It comprises 3,778 annotated frames extracted from 85 video clips recorded across three commercial pig farms (Durofarm1, Sunds1, Ikadan1), with 15 anatomically defined keypoints per pig instance annotated in COCO keypoint format. Three floor types are represented: cast iron slatted, concrete, and plastic slatted. The dataset includes a video-level stratified train/test split (68 train videos / 3,099 frames; 17 test videos / 679 frames) designed to prevent data leakage from near-duplicate frames. Five pose estimation architectures are benchmarked: SLEAP UNet (AP 0.951), ResNet-50 (AP 0.955), HRNet-W32 (AP 0.958), RTMPose-m (AP 0.935), and ViTPose-S (AP 0.960). Full documentation, keypoint schema, and benchmark results are provided in README.md.