A Task-Driven Feedback Imager with Uncertainty Driven Hybrid Control.

Area of Science:

• Computer Vision
• Machine Learning
• Control Systems

Background:

• Deep Neural Networks (DNNs) often exhibit overconfident outputs, limiting their use in critical feedback systems.
• Active sensor systems relying on DNNs require reliable uncertainty estimation for robust operation.

Purpose of the Study:

• To investigate uncertainty estimation in DNNs for closed-loop feedback smart cameras.
• To develop and evaluate an uncertainty-driven control strategy for enhancing feedback operations.

Main Methods:

• Estimated DNN uncertainty using sampling and non-sampling based methods.
• Proposed a closed-loop control system integrating uncertainty information.
• Implemented control modes prioritizing false positives, false negatives, and a hybrid approach.

Main Results:

• The hybrid uncertainty-driven control improved object detection and tracking accuracy by 1.1% on the CAMEL dataset.
• Action detection accuracy was enhanced by 1.4% using the hybrid approach.
• Demonstrated the effectiveness of uncertainty estimation in DNN feedback systems.

Conclusions:

• Uncertainty estimation is crucial for calibrating DNN outputs in feedback control.
• The proposed uncertainty-driven control strategies significantly enhance performance in computer vision tasks.
• This work paves the way for more reliable and accurate active sensor systems.