eCoMAE
An evidence-based masked pretraining method for multivariate time series that combines block-wise time-step masking with a Normal-Inverse-Gamma decoder. Through an observability-aware mechanism, it improves the reliability of uncertainty estimation under partially observable conditions.
Timeline (last 90 days)
eCoMAE is an evidence-masked pretraining method via observability alignment, used to improve the reliability of VRU motion prediction in out-of-distribution (OOD) scenarios.
In VRU intent prediction experiments on the IMPTC benchmark, eCoMAE improved the OOD accuracy of the strongest baseline from 0.7971 to 0.8068.
eCoMAE reduced the OOD AURC (area under the risk-coverage curve) from 0.0718 to 0.0648.
On the pedestrian trajectory forecasting task, eCoMAE achieved the best fine-tuned MSE on all four datasets: IMPTC, Waymo, nuScenes and inD.
eCoMAE combines block-wise temporal-step masking with a Normal-Inverse-Gamma decoder, and introduces three observability-aware mechanisms: uncertainty-weighted mean consistency, evidence suppression at masked positions, and representation consistency across random masks.
It has been argued that the effectiveness of eCoMAE in more complex dynamic scenarios (such as dense crowd interaction) still requires further validation.
All Facts (6)
eCoMAE is an evidence-masked pretraining method via observability alignment, used to improve the reliability of VRU motion prediction in out-of-distribution (OOD) scenarios.
50%UnverifiedIn VRU intent prediction experiments on the IMPTC benchmark, eCoMAE improved the OOD accuracy of the strongest baseline from 0.7971 to 0.8068.
50%UnverifiedeCoMAE reduced the OOD AURC (area under the risk-coverage curve) from 0.0718 to 0.0648.
50%UnverifiedOn the pedestrian trajectory forecasting task, eCoMAE achieved the best fine-tuned MSE on all four datasets: IMPTC, Waymo, nuScenes and inD.
50%UnverifiedeCoMAE combines block-wise temporal-step masking with a Normal-Inverse-Gamma decoder, and introduces three observability-aware mechanisms: uncertainty-weighted mean consistency, evidence suppression at masked positions, and representation consistency across random masks.
50%UnverifiedIt has been argued that the effectiveness of eCoMAE in more complex dynamic scenarios (such as dense crowd interaction) still requires further validation.
50%