ScNet: Scene-Consistency Network Learning for Multi-Agent Motion Forecasting

Predicting the motion of traffic agents is a fundamental challenge in autonomous driving, essential for safe and efficient ego-vehicle planning. Traditional methods typically focus on marginal forecasting, where the trajectory of each agent is predicted separately, leading to inconsistencies in scene-level predictions. To address this issue, we propose a scene-consistency network, named ScNet, which jointly predicts the trajectories of multiple agents in a single feedforward pass, ensuring consistency across all predictions. Our method leverages dual independently initialized student models that interact through cross-network contrastive learning at the global feature level, enhancing robustness and scene consistency in the learned representations. To further improve scene coherence, we incorporate a scene-guided strategy that refines these representations. Additionally, we employ a lightweight, anchor-free decoder that generates predictions for all agents, aligning the forecasts with real-world dynamics. Experiments show significant improvements in multi-world prediction metrics across complex environments. Code and models will be publicly available.