eCoMAE aligns evidential pretraining with observed time-series content
Original title: Observability-Aligned Evidential Masked Pretraining for Reliable VRU Motion Prediction
On IMPTC weather-shift tests, OOD accuracy rose from 0.7971 to 0.8068 and OOD AURC fell from 0.0718 to 0.0648.
The paper introduces eCoMAE, an evidential masked pretraining method for multivariate time series that ties predictive evidence to what is observed. It uses block-wise time-step masking, a Normal-Inverse-Gamma decoder, uncertainty-weighted consistency on jointly visible positions, evidence suppression on masked positions, and representation consistency across stochastic masks. On the IMPTC vulnerable-road-user intention benchmark under leave-one-environment-out weather shift, it improves the strongest baseline in out-of-distribution accuracy and AURC. Fine-tuning also reports the best mean squared error on pedestrian trajectory forecasting across IMPTC, Waymo, nuScenes, and inD.
- The target setting is partial observability and distribution shift for VRU prediction.
- OOD accuracy moved from 0.7971 to 0.8068 against the strongest baseline.
- OOD AURC moved from 0.0718 to 0.0648.
- Best fine-tuning MSE is reported on IMPTC, Waymo, nuScenes, and inD.
From the original article
1 min read at arxiv.orgVulnerable road user (VRU) prediction must operate under partial observability and distribution shift, so reliable uncertainty matters alongside accuracy.
By Shang Gao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.