Optimizing MPJPE promotes miscalibration in multi-hypothesis human pose lifting

Paweł A. Pierzchlewicz, Mohammad Bashiri, R. James Cotton, Fabian H. Sinz

Keywords: Pose estimation, lifting, multi-hypothesis, calibration, metrics

TL;DR: Optimizing MPJPE promotes miscalibration in multi-hypothesis human pose lifting

Abstract: Due to depth ambiguities and occlusions, lifting 2D poses to 3D is a highly ill-posed problem. Well-calibrated distributions of possible poses can make these ambiguities explicit and preserve the resulting uncertainty for downstream tasks, thus providing the necessary trustworthiness in safety-critical domains. This study shows that multi-hypothesis pose estimation methods produce miscalibrated distributions. We identify that miscalibration can be attributed to the optimization of mean per joint position error (MPJPE). In a series of simulations, we show that minimizing minMPJPE, the MPJPE of the best hypothesis, converges to the correct mean prediction. However, it fails to correctly capture the uncertainty, hence resulting in a miscalibrated distribution