MT-SRNet: A Transferable Multi-Task Super-Resolution Network for Pig Keypoint Detection, Segmentation, and Posture Estimation

Abstract

Robust visual analysis of pigs in densely housed groups is essential for various applications within digital phenotyping and Precision Livestock Farming, including animal counting, posture recognition, tracking, body dimension measurement, and welfare assessment. However, the transferability and scalability of existing methods across diverse farming environments remain significant challenges. In this study, we addressed these issues by proposing a modular design that decouples environmental variations from individual animal variability. At the framework-level, we introduced a Rotated Bounding Box (RBB) detector combined with an Auto-Visual Prompt strategy to effectively suppress background interference and achieve seamless scene adaptation. At the task-level, we developed a Multi-task Super-Resolution Network (MT-SRnet) capable of simultaneously predicting pig keypoints, masks, and postures from low-resolution inputs. Our experimental results demonstrate that MT-SRnet maintains high accuracy (Keypoints – 95.33 % mPCK; Mask – 95.53 % mIoU; posture – 89.72 % accuracy) while substantially reducing model complexity by over 30-fold. Moreover, the proposed approach achieves a substantial increase in inference speed (up to ∼17,000 pigs/s), highlighting its practical suitability for real-time monitoring under resource-limited conditions. More information available online: https://gitlab.kuleuven.be/m3-biores/public/m3pig.