Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

VerifIoU -- Robustness of Object Detection to Perturbations

Noémie Cohen, Mélanie Ducoffe, Ryma Boumazouza, Christophe Gabreau, Claire Pagetti, Xavier Pucel, Audrey Galametz

Abstract

We introduce a novel Interval Bound Propagation (IBP) approach for the formal verification of object detection models, specifically targeting the Intersection over Union (IoU) metric. The approach has been implemented in an open source code, named IBP IoU, compatible with popular abstract interpretation based verification tools. The resulting verifier is evaluated on landing approach runway detection and handwritten digit recognition case studies. Comparisons against a baseline (Vanilla IBP IoU) highlight the superior performance of IBP IoU in ensuring accuracy and stability, contributing to more secure and robust machine learning applications.

VerifIoU -- Robustness of Object Detection to Perturbations

Abstract

We introduce a novel Interval Bound Propagation (IBP) approach for the formal verification of object detection models, specifically targeting the Intersection over Union (IoU) metric. The approach has been implemented in an open source code, named IBP IoU, compatible with popular abstract interpretation based verification tools. The resulting verifier is evaluated on landing approach runway detection and handwritten digit recognition case studies. Comparisons against a baseline (Vanilla IBP IoU) highlight the superior performance of IBP IoU in ensuring accuracy and stability, contributing to more secure and robust machine learning applications.
Paper Structure (15 sections, 12 equations, 15 figures)

This paper contains 15 sections, 12 equations, 15 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Common concepts in object detection
  4. Robustness of object detection
  5. Perturbations applied to object detector
  6. Notations related to bounds
  7. Overview of the proposed approach
  8. Detailed description of the approach
  9. Vanilla_IoU - bounding the primitive operators
  10. Optimal_IoU extension - exact bounds
  11. Experiments
  12. Results
  13. Conclusion
  14. Appendices: Interval arithmetic operations
  15. Appendices: Partial derivatives

Figures (15)

  • Figure 1: Impact of input perturbations captured by the IoU.
  • Figure 2: Quantitative example of IoU
  • Figure 3: Flow chart illustrating the two-step approach to derive bounds for the IoU for a specific test image $s_0$ and perturbation domain $\Omega$.
  • Figure 4: IoU bound computation for a set of predicted bounding boxes
  • Figure 5: Equations for optimal_IoU
  • ...and 10 more figures

Theorems & Definitions (6)

  • Definition 1: Bounding box
  • Definition 2: Single class/single object detection model
  • Definition 3: Intersection over Union -- IoU
  • Definition 4: Perturbation domain
  • Definition 5: Bounds for predicted bounding boxes
  • Definition 6: Bounds for IoU