This paper introduces a framework for uncertainty quantification in pill recognition systems to address challenges in real-world cluttered environments. The proposed method integrates multi-scale feature extraction with evidential deep learning for explicit uncertainty decomposition and a post-hoc calibration module optimized for cluttered scenes.
Key findings
Identifies research gaps in existing pill recognition systems regarding uncertainty estimates and out-of-distribution detection.
Proposes a novel failure prediction metric considering aleatoric, epistemic, and detection uncertainties.
Plans to improve failure detection AUROC by 15-25% and provide reliable uncertainty estimates for cluttered scenarios.
Limitations & open questions
Potential excessive uncertainty estimates on visually similar pills.
Computational overhead from ensemble components.