Current motion forecasting approaches typically train a deep end-to-end model from the source domain data, and then apply it directly to target subjects. Despite promising results, they remain ...