• Introduces new metrics for transparent benchmarking of proposal efficiency. • Trigger-based framework reduces redundancy in behaviour proposal generation. • Enables robust detection of multi-class concurrent pig interactions. • Provides the PigInteraction dataset for multi-class social behaviour analysis. Understanding social interactions in group-housed pigs is critical for the understanding and improvement of animal welfare, health monitoring, and advancing precision livestock farming. Computer vision is emerging as an objective and accurate manner to detect, classify and quantify important behaviors of farm animals. However, a major challenge lies in accurately recognizing multiple behavior categories that often occur concurrently and overlap in both space and time within untrimmed video data. To address this, we propose a trigger-based approach for spatiotemporal behavioral proposal generation, which achieves 100% recall while significantly reducing redundant background proposals across five key behavior types: fighting, tail biting, displacement at the feeder, feeder-visiting, and drinker-visiting. Combining evolutionary parameter optimization with posture recognition, keypoint detection, and spatial constraints, our approach enables efficient and scalable detection even in densely populated animal housing environments. This method achieves 100% recall for five welfare-relevant behaviors while reducing redundant spatiotemporal proposals by ∼ 10×, enabling efficient analysis of long, dense farm recordings. A unified, modular trigger framework built over standard key-points and rotated bounding box that support both animal to animal and animal to object interactions. Using this approach, we also constructed a publicly available benchmark dataset with efficiency metrics, called PigInteraction, which features diverse and well-annotated interaction samples for downstream behavior analysis. The dataset is accessible at: https://gitlab.kuleuven.be/m3-biores/public/m3pig .
Liu et al. (Fri,) studied this question.