Maaninka VOS Dataset: Multi-Camera Video Object Segmentation Benchmark of Dairy Cows

A 1.5-hour, two-camera Video Object Segmentation (VOS) benchmark dataset for tracking 24 dairy cows in a research barn at the Natural Resources Institute Finland (Luke). Challenge: Evaluates long-term tracking, heavy occlusions, and multi-camera re-identification as cows move between two camera views. Annotations: Frame-by-frame instance segmentation masks with consistent IDs (1–24) created using a human-in-the-loop tool (SAM3 + XMem). Format: Videos are downsampled to 640×854 resolution at 10 FPS. Structure: Organized into left_cam and right_cam folders, each containing JPEGImages/ (JPG frames), Annotations/ (PNG mask files), raw video.mp4, and preview golden.mp4.