SPARC Enhances Motion Forecasting with Efficient Uncertainty Estimates

Human motion forecasting models are becoming more accurate, but their real-world deployment requires robust uncertainty estimates that are both structured and efficient. Traditional methods often involve computationally intensive stochastic inference or lack the ability to preserve trajectory covariance. This new framework, SPARC (Single-Pass Adaptive Risk Calibration), addresses these limitations by integrating a Bayesian last layer with conformal calibration. SPARC uses a deterministic neural network backbone to predict future motion, then applies a conjugate Bayesian layer to analytically derive an epistemic scale. This scale inflates a graph-temporal Gaussian covariance, maintaining its correlation structure, and is further calibrated using split conformal methods to produce valid prediction tubes. This innovative approach allows for the injection of feature-space epistemic uncertainty into trajectory densities without the need for Monte Carlo sampling. Evaluations across multiple datasets and protocols show SPARC outperforming existing deterministic, multimodal, and calibration baselines in terms of negative log-likelihood (NLL) and a combined accuracy-uncertainty criterion. It also maintains competitive point accuracy and offers an efficient way to monitor risk by separating high-error cases.