Algorithm 1: QCNet Training Procedure for Multi-Modal Trajectory Prediction

Input: Historical trajectories X, lane topology graph $\mathcal{G}$, ground-truth future     trajectories ${\mathbf{Y}}^{gt}$ (Let ${\mathbf{Y}}^{gt}={\left\{{\mathbf{y}}_i^{T_h+1:T_h+T_f}\right\}}_{i=1}^N$, denote the set of all     ground-truth future trajectories);   Output: Trained QCNet parameters $\theta$;