Motor ability-aware adaptive pose estimation with hierarchical uncertainty modeling and cross-ability learning

The current human pose estimation systems suffer from large decreases in accuracy as they are applied to people of different motor abilities. Current approaches to human pose estimation rely heavily upon a standardized anatomy that does not allow for the quantification of uncertainty. In this paper we propose an extensive framework for the human pose estimation problem that is composed of three sub-modules. The first module, called MADSRNet, dynamically alters the structure of skeletons to represent varying motor abilities by utilizing gated fusion and dynamic graph construction. The second module, called HUGPose, utilizes heteroscedastic regression to provide hierarchical uncertainty estimates at the joints, limbs, and body level. Finally, the third module, called CADCL, utilizes both contrastive learning and domain adversarial training to enable the transfer of knowledge across different motor abilities. We evaluate our proposed framework on the DiverseMotor-PE dataset and obtain state-of-the-art results (41.2 mm MPJPE), which represents a 14.7% improvement compared to the best baseline (ViTPose). The difference in mean per-level MPJPE for typical individuals and those with severe motor impairments is reduced from 12.1 mm to 10.3 mm. The expected calibration error of the uncertainty estimates is 1.58%. The motor ability aware design of our framework is able to make accurate predictions for humans regardless of their motor ability. This work contributes to the development of inclusive human pose estimation systems and promotes equitable access to healthcare and assistive technologies.