UncerTrans: uncertainty-aware temporal transformer for early action prediction

Early action prediction aims to predict complete action intentions by observing the initial stages of action execution, which is crucial for safety-critical applications such as human-robot collaboration and intelligent monitoring. Existing methods primarily focus on improving prediction accuracy but neglect the inherent uncertainty in early observations, leading to systems that cannot distinguish between reliable predictions and uncertain guesses. This paper proposes the UncerTrans framework, which combines Temporal Transformer with Monte Carlo Dropout to achieve accurate and trustworthy early action prediction. Temporal Transformer extracts discriminative features from extremely short observation sequences through hierarchical temporal attention mechanisms and temporal decay positional encoding. Monte Carlo Dropout generates prediction distributions through multiple random forward propagations during inference to quantify the model's epistemic uncertainty. An adaptive sampling strategy is designed to dynamically adjust sampling frequency based on initial uncertainty, balancing prediction quality and computational efficiency. Experiments on the EPIC-KITCHENS-100 dataset demonstrate that UncerTrans achieves 65.5% accuracy with only 10% observation ratio and an Expected Calibration Error of merely 0.089, significantly outperforming baseline methods. Through selective rejection of high-uncertainty predictions, the system can improve the accuracy of remaining predictions to 84.2%. The research demonstrates that effective uncertainty quantification relies on high-quality feature extraction, and the combination of both components enables early prediction systems to adopt differentiated strategies based on confidence levels, providing a technical foundation for practical deployment.