GANPose: Pose estimation of grouped pigs using a generative adversarial network

Pose estimation as an effective method for behaviour detection has attracted wide attention in the field of animal health and welfare detection in recent years. Consequently, we can monitor individual pig behaviour changes over time, which will assist in the early detection of disease, allowing for earlier and more effective intervention. However, a real farming scenario with occlusions and opaque pig clustering may be a challenging computer vision problem. Inspired by human vision, we propose GANPose, a grouped pig pose estimation model based on a generative adversarial network (GAN). This model uses the adversarial game to induce the model to learn prior information about the pig's body structure and output a pose with a reasonable skeleton structure. In addition, we also established a new pig pose estimation dataset, which not only included the basic back keypoints but also small keypoints, such as leg, elbow, and tail, as well as challenging occluded keypoints. This work is the study to tackle the occlusion problem, with a finely tuned GAN structure. Experiments and quantitative results on our dataset show that GANPose can detect keypoints of herd pigs with 74.09 % mAP, and the accuracy of the percentage of correct keypoint (PCK) is 11.37 % higher than the benchmark for occluded keypoints, achieving excellent performance in a real pigsty environment.