Thermal Image Dataset for Human Pose Keypoint Detection and Action Recognition
Description
This dataset consists of 2,500 thermal images at 720×576 resolution, annotated for human pose keypoint detection using the YOLO-Pose format with 17 COCO-style human keypoints. Action dataset consists of 6 classes: - walking: 30797 frames 20.5 min - sitting: 27487 frames 18.3 min - standing: 20511 frames 13.7 min - lying: 14275 frames 9.5 min - calling: 13050 frames 8.7 min - drinking: 6369 frames 4.2 min Action annotation format: [frame, walking, sitting, standing, drinking, calling, lying, kp1_x, kp1_y, kp1_v, ... 17*3 keypoints, ... , kp17_x, kp17_y, kp17_v] Keypoints are converter to bbox-relative coordinates. Use the script "Action-Recognition/2play_dataset.py" for visualization. Use the script "Action-Recognition/thermal-lstm-train-model-in-loop.ipynb" to setup and train Keras models.