核心概念
IBP IoU introduces a novel approach for formal verification of object detection models, focusing on stability and accuracy.
摘要
The IBP IoU approach aims to verify the stability of object detection models using the Intersection over Union (IoU) metric. By implementing perturbations and interval bound propagation, the method ensures that the model remains stable under various conditions. The study evaluates the performance on landing approach runway detection and handwritten digit recognition, showcasing superior accuracy and stability compared to baseline methods. The research addresses the critical need for formal verification in machine learning applications, emphasizing correctness and robustness.
統計資料
The experiments were parallelized over 20 workers on a Linux machine with an Intel Xeon processor.
MNIST dataset used for handwritten digit localization, LARD dataset for runway detection during landing.
Perturbation types include white noise, brightness, and contrast with varying parameters.
Optimal IoU outperforms Vanilla IoU in certifying box stability across different datasets and perturbations.
引述
"We propose a two-step approach using classical verification tools to obtain reachable outputs."
"Our method ensures stability against local perturbations by bounding the challenging IoU function."
"Optimal IoU extension provides exact bounds for ensuring stability of object detection models."