Training-Free Uncertainty Estimation for Dense Regression: Sensitivity as a Surrogate

Uncertainty estimation is an essential step in the evaluation of robustness for deep learning models computer vision, especially when applied risk-sensitive areas. However, most state-of-the-art either fail to obtain uncertainty or need significant modification (e.g., formulating a proper Bayesian treatment) it. Most previous methods are not able take arbitrary model off shelf and generate without retraining redesigning To address this gap, we perform systematic exploration into training-free dense regression, unrecognized yet important problem, provide theoretical construction justifying such estimations. We propose three simple scalable analyze variance outputs from trained network under tolerable perturbations: infer-transformation, infer-noise, infer-dropout. They operate solely during inference, re-train, re-design, fine-tune models, as typically required by methods. Surprisingly, even involving perturbations training, our produce comparable better compared training-required Code available at https://github.com/lumi9587/train-free-uncertainty.