Mouse Pose Estimation Using Synthetic Datasets Replicating Experimental Environments

Quantitative measurement of animal behavior using machine learning techniques has gained significant attention, and stable animal pose estimation is crucial for this purpose. While various methodologies for pose estimation using deep learning have been proposed in ideal experimental settings, few have addressed challenges in real-world environments with high occlusion. In this study, we aimed to achieve stable animal pose estimation under these conditions. Synthetic data generation has emerged as a viable solution that offers a cost-effective approach compared with real data annotation. Our approach involves creating synthetic datasets tailored to specific experimental scenarios, such as the foot fault test. We developed 3D models of mice and experimental setups, which enabled us to render synthetic datasets with diverse environmental factors. We evaluated the effectiveness of pose estimation using these synthetic datasets, achieving promising results that demonstrate the potential for real-world applications in animal behavior research. Further advancements include 2D and 3D pose estimation techniques validated through comprehensive testing and evaluation.