A new arXiv preprint argues that despite advances in sub-pixel decoding, human pose estimation (HPE) architectures remain limited by deterministic coordinate extraction from discrete heatmaps. The authors state that this approach systematically fails in scenarios with dense occlusions, where keypoint locations are ambiguous or hidden.
The proposed remedy is a probabilistic formulation that models keypoint locations as a mixture of Gaussian and uniform distributions. This mixture is intended to capture both precise localization and the uncertainty introduced by occlusions, moving away from a single deterministic output.
The abstract does not provide experimental details or quantitative results, so the claimed benefits are not yet verified. Still, the framing points to a clear limitation in current HPE pipelines and offers a probabilistic direction for addressing it.