THOMAS: Trajectory Heatmap Output with learned Multi-Agent Sampling
This addresses the problem of predicting safe and coherent future paths for multiple interacting agents (e.g., vehicles or pedestrians), which is crucial for autonomous systems, though it appears incremental as it builds on existing multi-agent prediction methods.
The paper tackles multi-agent trajectory prediction by proposing THOMAS, a framework that jointly predicts consistent multi-modal trajectories for multiple agents, achieving first place on the Interaction multi-agent prediction challenge leaderboard.
In this paper, we propose THOMAS, a joint multi-agent trajectory prediction framework allowing for an efficient and consistent prediction of multi-agent multi-modal trajectories. We present a unified model architecture for simultaneous agent future heatmap estimation, in which we leverage hierarchical and sparse image generation for fast and memory-efficient inference. We propose a learnable trajectory recombination model that takes as input a set of predicted trajectories for each agent and outputs its consistent reordered recombination. This recombination module is able to realign the initially independent modalities so that they do no collide and are coherent with each other. We report our results on the Interaction multi-agent prediction challenge and rank $1^{st}$ on the online test leaderboard.