Patient Vital Extraction from Images of ICU Patient Monitors

It contains 402 photographic images of ICU patient monitor displays curated to benchmark screen localization and vital sign extraction across zero-shot segmentation (SAM), supervised object detection (YOLOv8), and Vision-Language Models (Molmo2-4B). The collection comprises 198 web-sourced clinical monitor images and phone captured Dräger Infinity Delta monitor across varying distances, optical skews, and lighting/glare conditions.

Annotations provide screen bounding boxes and scalar values for Heart Rate (HR, green digits) and Blood Oxygen Saturation (SpO2, blue and white digits). The data is partitioned into training (140), validation (30), and testing (30) splits.