3DB: A Framework for Debugging Computer Vision Models

By Guillaume Leclerc et al

Read the original document by opening this link in a new tab.

Table of Contents

1 Introduction
2 Designing 3DB
3 Debugging and analyzing models with 3DB
3.1 Sensitivity to image backgrounds
3.2 Texture-shape bias
3.3 Orientation and scale dependence

Summary

The paper introduces 3DB, a framework for testing and debugging computer vision models using photorealistic simulation. It allows users to identify vulnerabilities in computer vision systems and gain insights into model decision-making. 3DB captures and generalizes robustness analyses, enabling users to study the interplay between different factors. The paper discusses the design principles behind 3DB, its capabilities, and workflow. Case studies are presented to demonstrate how 3DB can be used to analyze and debug vision models, focusing on sensitivity to image backgrounds, texture-shape bias, and orientation-scale dependence.