Enhancing Trajectory Prediction using Sparse Outputs: Application to Team Sports