Understanding Uncertainty Maps in Vision with Statistical Testing.

Area of Science:

• Computer Vision
• Machine Learning
• Statistical Modeling

Background:

• Quantitative uncertainty descriptions are crucial for deep neural network (DNN) models in machine learning and computer vision.
• Existing methods for statistical testing of uncertainties from overparameterized models are limited, especially for high-resolution images.
• Evaluating the uncertainty behavior of different models with similar accuracy profiles remains a challenge.

Purpose of the Study:

• To develop an efficient framework for hypothesis testing of uncertainty maps generated by DNNs.
• To address the lack of statistical testing capabilities for model uncertainties in computer vision applications.
• To enable statistically significant comparisons of uncertainty behaviors between different models.

Main Methods:

• Revisiting Random Field Theory (RFT) principles.
• Integrating RFT with deep neural network (DNN) tools to overcome computational challenges.
• Developing hypothesis testing frameworks specifically for image-based uncertainties.

Main Results:

• Demonstrated an efficient framework for hypothesis testing of uncertainty maps from DNNs.
• Showcased the ability to perform statistically significant comparisons of model uncertainty behaviors.
• Validated the framework's viability across various computer vision experiments.

Conclusions:

• The proposed RFT-enhanced DNN framework provides novel hypothesis testing capabilities for image uncertainties.
• This approach facilitates statistically rigorous evaluation and comparison of uncertainty quantification in deep learning models.
• The method is applicable to critical applications requiring reliable uncertainty assessment in computer vision.