Improve Pose Estimation Model Performance with Unlabeled Data

The infant's spontaneous motricity is a fundamental element in the clinical evaluation of neonates. Therefore, the pediatrician values any modification in the neonate's motor pattern for diagnosing infectious, genetic, metabolic, or neuromuscular disease. Automatization of spontaneous motricity analysis would be of great help in diagnosing and treating patients earlier. However, as the patterns are very numerous, it would be resources consuming to use labeled data. We aim to eliminate the necessity for a tedious and costly manual annotation. Our work presents a new infant motor skills dataset and describes methods to improve the automatic 2D pose estimation with AlphaPose from RGB videos. We evaluate the results using PCKh and a clinician's assessment. The results show that our methods significantly increase the AlphaPose performances on our dataset without needing labeled data. We tested the relevance of the posture estimation obtained with our models by training a classification model that separates a motricity of a control group from the motricity of children with Spinal Muscular Atrophy; our model achieves encouraging results for further research with more complex models.