Caged Monkey Dataset: A New Benchmark for Caged Monkey Pose Estimation

Abstract

AbstractAutomatic monkey pose estimation is of great potential for quantitative behavior analysis of monkeys, which provides indispensable information for the studies of drug safety assessments and medical trials. With the development of deep learning, the performance of human pose estimation has been greatly improved, however, the study of monkey pose estimation is rare and the robustness of performance is unsatisfactory due to the lack of data and the variations of monkey poses. In this work, we propose a complete solution to address these problems in terms of data and methodology. For data, we collect a comprehensive Caged Monkey Dataset with 6021 samples, with labeled poses. For methodology, we propose a Mask Guided Attention Network (MGAN) to focus on the foreground target automatically so as to locate the occluded keypoints precisely and deal with the complex monkey postures. The proposed method is evaluated on the collected monkey dataset, achieving 79.2 Average Precision (AP), which is 6.1 improvement over the baseline method and is comparable with the state-of-the-art performance of human pose estimation. We hope our attempt on caged monkey pose estimation will serve as a regular configuration in drug safety assessments and medical trials in the future.KeywordsCaged monkey datasetMonkey pose estimationMask guided attention networkDrug safety assessments