Uncertainty Estimation for Safety-critical Scene Segmentation via Fine-grained Reward Maximization

• DOI: 10.48550/arxiv.2311.02719
• Publication Date: 2023-01-01
• Authors: Hongzheng Yang; Cheng Chen; Yueyao Chen; M. W. Scheppach; Hon Chi Yip; Qi Dou
• Keywords: Maximization; Margin (machine learning)

Uncertainty estimation plays an important role for future reliable deployment of deep segmentation models in safety-critical scenarios such as medical applications. However, existing methods uncertainty have been limited by the lack explicit guidance calibrating prediction risk and model confidence. In this work, we propose a novel fine-grained reward maximization (FGRM) framework, to address directly utilizing metric related function with reinforcement learning based tuning algorithm. This would benefit through direct optimization calibration. Specifically, our method designs new using calibration metric, which is maximized fine-tune evidential pre-trained risk. Importantly, innovate effective parameter update scheme, imposes reward-weighting each network according importance quantified fisher information matrix. To best knowledge, first work exploring vision tasks. The effectiveness demonstrated on two large surgical scene datasets under different settings. With real-time one forward pass at inference, outperforms state-of-the-art clear margin all metrics estimation, while maintaining high task accuracy results. Code available \url{https://github.com/med-air/FGRM}.