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

Accurate detection of sow postures provides valuable insights into animal activity, which is a key indicator for health, welfare, and productivity. While activity monitoring using wearable sensors has become a standard approach for certain livestock species like dairy cattle, its application in pig production remains limited due to practical challenges. Advances in computer vision (CV) and artificial intelligence (AI) offer an alternative, enabling non-invasive monitoring of pig behavior. However, the development of robust CV models requires large and diverse datasets that capture the variability of housing environments and visual conditions. To address this limitation, we present SowPostureDS, a dataset comprising 14,400 annotated images of sows from different farrowing systems. Images were extracted from long-term video recordings and manually annotated into four posture classes following a consistent scheme. The dataset reflects heterogeneity in housing design, lighting conditions, and the presence of piglets, providing a valuable resource for implementing CV models with improved robustness across diverse environments. A use case demonstrates the dataset's applicability, showing that models trained on SowPostureDS can achieve high accuracy and can be effectively adapted to unseen environments using transfer learning with minimal additional data.