Enforcing label consistency and lowering labelling time in infants’ pose data via semi-automatic annotation

Infants’ spontaneous movements provide clinically relevant information about neurodevelopment, but their assessment still relies on qualitative visual inspection, which is prone to variability. Video-based systems with automatic pose estimation algorithms have been proposed to address this, yet annotating data to train these algorithms is time-consuming and prone to intra- and inter-annotator variability. To improve the efficiency and consistency of annotation, semi-automatic annotation has been proposed in the broader human pose estimation literature. To assess the extent to which semi-automatic annotation may also be beneficial for infants’ pose estimation, we benchmarked six state-of-the-art adult-trained models on a new dataset comprising 46 videos of preterm infants recorded in a neonatal unit. The predicted joints from the best-performing model were presented to human annotators for optional refinement, who were also asked to label the same joints manually for comparison. ViTPose and Sapiens achieved the highest joint localisation accuracy, with ViTPose requiring markedly lower computational cost. Semi-automatic annotation reduced inter-annotator variability by 7.82% and decreased the time to label a frame by 35.9% compared to full manual labelling. These findings support the use of semi-automatic annotation as an effective strategy to enhance the labelling process of infant pose estimation data from real clinical settings.