A Verification Platform for Human Pose Estimation Models

Human pose estimation (HPE) from images and videos is an essential part of motion analysis and is increasingly used in various fields, ranging from Healthcare to Industry 5.0. Nevertheless, such CNN-based model has an intrinsic inaccuracy, which sources include insufficient CNN training, sensor low-quality, inadequate computational capability for the inference phase, or heavy occlusions in the real scene. Several techniques are continuously proposed to increase the HPE accuracy. They focus on more advanced convolutional neural networks (CNN), smarter training, or denoising and completion procedures. State-of-the-art methods to measure the system accuracy usually rely on testsets that differ from the specific conditions faced in the final real-world deployment and on solution-oriented evaluation metrics. This makes the evaluation of the accuracy of such platforms in real application scenarios imprecise and biased. To solve this problem, we propose an automatic verification platform in which the different types of inaccuracy sources are represented by error models and implemented by software mutants. The platform injects the mutants to simulate the inaccuracy and measures, through a large set of key indicators and statistics, the HPE software robustness. We evaluated our verification methodology on a popular HPE dataset with a state-of-the-art pose estimator.