THOMAS: Trajectory Heatmap Output with learned Multi-Agent Sampling

TL;DR

THOMAS is a novel multi-agent trajectory prediction framework that efficiently generates consistent multi-modal trajectories using hierarchical heatmaps and a learnable recombination module, achieving top performance on a benchmark.

Contribution

The paper introduces a unified model architecture with hierarchical heatmap estimation and a learnable recombination module for coherent multi-agent trajectory prediction.

Findings

Achieved 1st place on the Interaction multi-agent prediction challenge leaderboard.

Developed a memory-efficient, fast inference method for multi-agent trajectory prediction.

Ensured trajectory coherence and collision avoidance through a learnable recombination process.

Abstract

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.