Résumé
Reliable uncertainty estimation is critical for deploying 2D human pose estimation in safety-critical settings. We propose an evidential deep learning framework that jointly estimates 2D joint locations and their associated aleatoric and epistemic uncertainty. Our approach adopts a two-stage architecture: an evidential network first predicts uncertainty-aware joint heatmaps, which are subsequently refined by an evidential deformable decoder to obtain subpixel joint localization. This design provides a probabilistic representation of joint coordinates while avoiding Monte Carlo sampling and model ensembling. Experiments on standard benchmarks show that our method achieves competitive pose estimation accuracy while producing well-calibrated uncertainty estimates, outperforming existing deep evidential regression approaches for human pose estimation.