University Honors

Cerebral palsy is a neurological disorder characterized by abnormal, spontaneous limb movements in infants. Early detection is crucial for timely treatment, but such assessments are challenging and rigorous, requiring extensive training to objectively identify abnormal movement patterns. Pose estimation models enable automated, markerless joint tracking, which can be used to detect and quantify movements patterns. However, most models are built using adult subjects, and those for infants require labeled training data, which is very laborious to generate and difficult to access. To address these challenges, we built an adaptive method called SHIFT: Leveragine SyntHetic Adult Datasets for InFanT Pose Estimation. SHIFT incorporates innovative techniques like the mean teacher framework, an infant manifold pose prior, and a novel visibility consistency module to effectively adapt a pre-trained 2D adult pose estimation model to infants. We performed extensive experiments on multiple pose estimation benchmarks, including unsupervised domain adaptation and supervised methods. We demonstrate that SHIFT outperforms unsupervised domain adaptation methods by an increase of 5% (51% vs 56%, respectively) and supervised methods by an increase of 16% (68% vs 84%, respectively) in prediction accuracy. Building on this foundation, we con duct additional ablation studies to assess the optimal weighting of the manifold pose prior and keypoint segmentation modules, ensuring that anatomical constraints enhance rather than distort predictions during the adaptation process. We now propose SHIFT as an infant pose estimation method independent of labeled target domain data, which we hope will enable efficient assessment of disorders such as cerebral palsy using computer vision.