SowPostureDS: A Multi-Class Image Dataset for YOLO-Based Detection of Sow Postures in diverse Farrowing Systems

Area of Science:

• Animal Science
• Computer Vision
• Artificial Intelligence

Background:

• Sow activity monitoring is vital for health, welfare, and productivity.
• Wearable sensors are common in cattle but limited in pigs due to practical challenges.
• Computer vision (CV) and artificial intelligence (AI) offer non-invasive alternatives for pig behavior monitoring.

Purpose of the Study:

• To address the need for diverse datasets for robust CV models in pig production.
• To introduce SowPostureDS, a comprehensive dataset for sow posture analysis.
• To facilitate the development of accurate and adaptable CV models for sow behavior.

Main Methods:

• Collected 14,400 images of sows across different farrowing systems.
• Manually annotated images into four distinct posture classes.
• Ensured dataset heterogeneity in housing, lighting, and piglet presence.

Main Results:

• SowPostureDS provides a diverse resource for CV model development.
• Models trained on SowPostureDS achieve high accuracy in posture detection.
• Transfer learning with minimal data effectively adapts models to new environments.

Conclusions:

• SowPostureDS significantly advances non-invasive sow behavior monitoring.
• The dataset supports the creation of robust CV models for diverse agricultural settings.
• This resource enhances precision livestock farming for improved pig welfare and productivity.