Heteroscedastic Diffusion for Multi-Agent Trajectory Modeling.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multi-agent trajectory modeling often prioritizes forecasting over completion, crucial for data correction.
• Current methods lack state-wise heteroscedastic uncertainty and per-scene error probability estimates.
• This limits the ranking and reliability of predictions in real-world applications.

Purpose of the Study:

• To develop a unified diffusion model for trajectory completion with state-wise uncertainty estimation.
• To introduce a faster baseline model (U2Diff) for efficient trajectory modeling.
• To enable reliable ranking of generated trajectories using error probability estimation.

Main Methods:

• U2Diffine augments denoising loss with negative log-likelihood of predicted noise.
• Latent space uncertainty is propagated to real state space via first-order Taylor approximation.
• A Rank Neural Network (RankNN) is integrated for post-processing error probability estimation.

Main Results:

• U2Diffine achieves state-of-the-art performance in trajectory completion and forecasting.
• The proposed method provides accurate state-wise heteroscedastic uncertainty estimates.
• RankNN demonstrates strong correlation between estimated and ground truth errors, enabling effective prediction ranking.

Conclusions:

• The developed diffusion model effectively addresses trajectory completion and uncertainty estimation challenges.
• The integration of uncertainty and error probability estimation enhances the reliability of multi-agent trajectory predictions.
• The method shows significant improvements across diverse sports datasets, validating its practical applicability.