DOI: 10.11118/ecos2026_035

A UNIFIED FORMAT FOR AGGREGATING PRE-TRAINED POSE ESTIMATORS INTO CUSTOM HUMAN-SKELETON DATASETS

Damian Sova¹, Oldrich Trenz¹

¹ Department of Informatics, Faculty of Business and Economics, Mendel University in Brno, Zemědělská 1, 613 00 Brno, Czech Republic

Human pose estimation works well on benchmark datasets but loses accuracy once a model is deployed on real footage with varying distances, crowds and occlusion. Building a labelled dataset for each new deployment is expensive, so this work asks whether existing pre-trained detectors can be combined to generate such a dataset cheaply, without retraining any of them. The proposed Unified Format runs three public detectors (PoseNet, MoveNet, MMPose), maps their native keypoints into one common 27-keypoint skeleton, matches detected people across models and fuses them by weighted average. The method was evaluated on the COCO WholeBody validation subset (2,693 images with humans, 11,004 ground-truth instances) using OKS, average percentage error and mean squared error. The fused skeleton provides usable positional localisation (APE 3.3%, MSE 1,213.84) and predicts slightly more pose instances than MMPose alone (70.39% vs 69.84%), but it remains weaker than MMPose on the main accuracy-oriented metrics (OKS 0.23 vs 0.28; correct keypoints 37.93% vs 43.62%). The result is an honest, reproducible tool chain rather than a new state of the art.

Keywords: human pose estimation, neural networks, dataset generation, data fusion, COCO WholeBody, OKS

pages: 323-329, online: 2026