Fig. 1.
The proposed GaitForeMer framework for motor impairment severity estimation. A Transformer model based on body skeletons is first pre-trained on a large dataset jointly for human motion prediction and activity prediction. The model is subsequently fine-tuned on the task of MDS-UPDRS gait score prediction using extracted 3D skeleton sequences from clinical study participants using VIBE (Video Inference for Body pose and shape Estimation) [13].