Reproduce “Annealed Winner-Takes-All for Motion Forecasting” on OpenResearch

About this paper

Yihong Xu, Victor Letzelter, Mickaël Chen, Éloi Zablocki, Matthieu Cord

This work introduces an annealed Winner-Takes-All (aWTA) loss for motion forecasting models, which consistently improves prediction accuracy and training stability compared to the standard WTA loss. The method enables state-of-the-art models to achieve better performance with a minimal number of hypotheses (e.g., 6) for both training and inference, eliminating the need for post-selection steps like NMS.

Key insights

• Annealed optimization, through aWTA, effectively mitigates mode collapse and training instabilities inherent in standard WTA, allowing models to learn a more diverse and accurate set of multi-modal trajectories.
• The gradual annealing of the temperature parameter enables robust training with a minimal number of hypotheses, significantly streamlining the inference pipeline by removing the need for post-selection.
• The aWTA approach demonstrates robust performance across a broad range of initial temperature and decay rate hyperparameters, simplifying its adoption and integration into existing architectures.

Summary by alphaXiv.