On complementing end-to-end human motion predictors with planning