Abstract
Pig instance segmentation is a critical component of smart pig farming, serving as the basis for advanced applications such as health monitoring and weight estimation. However, existing methods typically rely on large volumes of precisely labeled mask data, which are both difficult and costly to obtain, thereby limiting their scalability in real-world farming environments. To address this challenge, this paper proposes a novel approach that leverages simpler box annotations as supervisory information to train a pig instance segmentation network. In contrast to traditional methods, which depend on expensive mask annotations, our approach adopts a weakly supervised learning paradigm that reduces annotation cost. Specifically, we enhance the loss function of an existing weakly supervised instance segmentation model to better align with the requirements of pig instance segmentation. We conduct extensive experiments to compare the performance of the proposed method that only uses box annotations, with that of five fully supervised models requiring mask annotations and two weakly supervised baselines. Experimental results demonstrate that our method outperforms all existing weakly supervised approaches and three out of five fully supervised models. Moreover, compared with fully supervised methods, our approach exhibits only a 3% performance gap in mask prediction. Given that annotating a box takes merely 26 seconds, whereas annotating a mask requires 94 seconds, this minor accuracy trade-off is practically negligible. These findings highlight the value of employing box annotations for pig instance segmentation, offering a more cost-effective and scalable alternative without compromising performance. Our work not only advances the field of pig instance segmentation but also provides a viable pathway to deploy smart farming technologies in resource-limited settings, thereby contributing to more efficient and sustainable agricultural practices.