Structured Differential Learning for Automatic Threshold Setting

TL;DR

This paper presents a heuristic method for automatically tuning parameters in rule-based computer vision systems, enabling gradient-based optimization while maintaining system flexibility, demonstrated on an automotive headlight controller.

Contribution

Introduces a novel heuristic technique for automatic parameter tuning in rule-based vision systems using approximate gradient descent, a previously unavailable efficient method.

Findings

Successfully tuned over 100 parameters with minimal data

Outperformed manual tuning in automotive headlight control

Demonstrated practical utility in real-world videos

Abstract

We introduce a technique that can automatically tune the parameters of a rule-based computer vision system comprised of thresholds, combinational logic, and time constants. This lets us retain the flexibility and perspicacity of a conventionally structured system while allowing us to perform approximate gradient descent using labeled data. While this is only a heuristic procedure, as far as we are aware there is no other efficient technique for tuning such systems. We describe the components of the system and the associated supervised learning mechanism. We also demonstrate the utility of the algorithm by comparing its performance versus hand tuning for an automotive headlight controller. Despite having over 100 parameters, the method is able to profitably adjust the system values given just the desired output for a number of videos.