Multi-pig Pose Estimation Using DeepLabCut

Abstract

Pose estimation towards providing the assessments of animal health and welfare monitoring has strongly gained interest in the last few years. However, it is a challenging computer vision problem as the frequent interaction causes occlusions the association of detected key-points to the correct individuals. Deep Learning (DL) offers major advances in the field of pose estimation. In this paper, we investigated the possibility of using a famous open-source DL-based toolbox, DeepLabCut [1], for the specific pig pose estimation task. We predicted the body part of each individual pig from only input images or video sequences directly with no adaptations to the application setting. We used a real dataset which contains 2000 annotated images with 24,842 individually annotated pigs from 17 different locations and light conditions. The experimental results demonstrated that we can achieve a small root mean square error between the manual and predicted labels (10.1) when detecting pigs in environments previously seen by a DL model during training. To evaluate the robustness of the trained model, it is also tested on environments and lighting conditions unseen in the training set, where it achieves 12.0 root mean square error.