Annotated Image Dataset for Computer-Vision-Based Tear Staining Assessment in Pigs

Description This dataset contains images of pig heads acquired using a dedicated tear-staining sensor installed at the bleeding line of a commercial slaughterhouse. The sensor was developed within the aWISH project and uses two cameras to record both sides of pigs passing the bleeding line. A protective panel separates the imaging hardware from the animals and protects the system from the slaughterhouse environment, while additional lighting is used to improve image quality. The published images were extracted from recordings collected under these commercial conditions. Three complementary annotation types are provided to support different stages of an automated tear-staining analysis pipeline: pig-head detection, head-pose estimation using keypoints, and segmentation of tear-staining and eye-surface regions. Head detection annotations Bounding-box annotations identify the visible pig head in each image. The annotations are provided in YOLO-compatible format: <class_id> <x_center> <y_center> <width> <height> All bounding-box coordinates are normalized to the image dimensions. Class ID Class 0 pig_head Head pose / keypoint annotations Head orientation is represented using four anatomical landmarks: right eye left eye nose forehead Each pig-head annotation contains four keypoints using a kpt_shape of [4,3]. Each keypoint is represented by normalized x and y coordinates and a visibility value. Class ID Class 0 head_skeleton Segmentation annotations Polygon annotations are provided for the tear-staining and visible eye-surface regions. Class ID Class 0 tear_staining 1 eye_surface Dataset characteristics Number of published images: 650 Image format: PNG Image resolutions: 1920 × 1080 and 2560 × 1440 Annotation format: YOLO-compatible text files Annotation types: bounding boxes, keypoints and segmentation polygons Known limitations The dataset originates from a specific sensor installation and commercial slaughterhouse environment. Camera geometry, illumination, background, occlusion, contamination and animal orientation may therefore differ from other recording systems or slaughterhouse conditions. Tear-staining boundaries can be visually ambiguous, particularly for faint staining or images collected under difficult conditions. Annotation uncertainty should therefore be considered when interpreting segmentation performance.