anilbhujel/Gilt_posture_dataset

Gilt Posture Recognition Dataset

• Each RGB image has a matching depth image (same filename, .png extension).
• YOLO-format label files correspond to each image.

🐷 Annotated Postures

Five postures are labeled using YOLO bounding boxes:

Class Name	Class ID
feeding	0
lateral_lying	1
sitting	2
standing	3
sternal_lying	4

📊 Class Distribution

Below is a histogram showing the distribution of posture classes across the dataset:

Dataset Description

The total dataset is split randomly into training, validation, and testing sets (0.75:0.15:0.1). The filename of each image and corresponding labels are assigned with date and time of image captured prefixed by pen and camera identity (p1c1_20250108_080409.png == image of pen1 camera1 captured on January 08, 2025 at 08:04:09 o'clock)

• The Color folder contains the color images and corresponding labels.
• Depth folder contains the height information of scene from the floor in mm unit and saved as uint16 format.
• RGBD folder contains the combined pairs of color and depth images. The normalized height information is added as 4th channel (RGBA).
• Each folder contains a labels folder for the corresponding labeling information

🧠 Use Cases

• Animal behavior monitoring
• Multimodal object detection (RGB + Depth)
• Precision livestock farming

License

The author has granted permission to download, use and redistribute this dataset only for research purposes.

Citation

Please cite as Bhujel A. et al. (2025). A Computer Vision dataset for Gilts' daily activity monitoring and tracking.

Contact

For questions or collaborations, feel free to reach out at [email protected]