UncerTrans: uncertainty-aware temporal transformer for early action prediction.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Early action prediction is vital for safety-critical systems like human-robot collaboration.
• Current methods often overlook prediction uncertainty, limiting trust and reliability.
• Distinguishing confident predictions from uncertain guesses remains a challenge.

Purpose of the Study:

• To develop a framework for accurate and trustworthy early action prediction.
• To quantify and manage the uncertainty inherent in predicting actions from limited initial observations.
• To enhance the practical deployment of action prediction systems in real-world scenarios.

Main Methods:

• Proposed the UncerTrans framework integrating Temporal Transformer and Monte Carlo Dropout.
• Temporal Transformer utilized hierarchical temporal attention and decay positional encoding for feature extraction from short sequences.
• Monte Carlo Dropout quantified epistemic uncertainty via multiple inference passes; adaptive sampling optimized efficiency.

Main Results:

• UncerTrans achieved 65.5% accuracy with only 10% observation data on EPIC-KITCHENS-100.
• Demonstrated low Expected Calibration Error (0.089), indicating reliable uncertainty quantification.
• Selective rejection of uncertain predictions improved accuracy to 84.2%.

Conclusions:

• Effective uncertainty quantification necessitates high-quality feature extraction.
• Combining robust feature extraction with uncertainty estimation enables confidence-based differentiated strategies.
• The UncerTrans framework provides a foundation for deploying reliable early action prediction systems.