OpenThermalPose and OpenThermalPose2 Datasets Card

Dataset Summary

This dataset card describes OpenThermalPose and its extension, OpenThermalPose2, two open-source datasets for thermal human pose estimation. These datasets contain annotated thermal images of humans performing various activities, including fitness exercises, multiple-person interactions, and walking in diverse outdoor and indoor environments. Annotations include bounding boxes and 17 keypoints per human instance, consistent with the MS COCO Keypoint dataset. Pre-trained YOLOv8-pose and YOLO11-pose models are also provided as baselines.

OpenThermalPose

• Images: 6,090
• Subjects: 31
• Annotated Instances: 14,315
• Annotations: Bounding boxes, 17 keypoints
• Activities: Fitness exercises, multiple-person activities, outdoor walking
• Conditions: Varied weather conditions and locations
• Pre-trained Models: Link to Pre-trained Models

OpenThermalPose2

• Images: 11,391
• Subjects: 170
• Annotated Instances: 21,125
• Annotations: Bounding boxes, 17 keypoints
• Activities: Fitness exercises, multiple-person activities, indoor sitting, outdoor walking
• Conditions: Varied weather conditions and locations, indoor and outdoor settings
• Pre-trained Models: Link to Pre-trained Models

Dataset Examples

(Images showcasing various activities and settings are included here. Refer to the original text for image links.)

Model Baselines

YOLOv8-pose and YOLO11-pose (nano, small, medium, large, and x-large) models were trained and evaluated on both datasets. Further details on training and evaluation can be found in the provided preprints. The models utilize the Ultralytics library.

Additional Information

• Preprint (OpenThermalPose): TechRxiv Link
• Preprint (OpenThermalPose2): TechRxiv Link
• Ultralytics Documentation: Link to Ultralytics Docs