Explaining Bounding Boxes in Deep Object Detectors Using Post Hoc Methods for Autonomous Driving Systems.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Deep learning models are crucial for autonomous driving perception tasks like object detection.
• The 'black-box' nature of these models necessitates explainability for safety and reliability.
• Existing explainable AI (XAI) techniques require thorough evaluation in the complex autonomous driving context.

Purpose of the Study:

• To explore and compare explainable AI (XAI) techniques for object detection in autonomous driving.
• To evaluate gradient-based and perturbation-based saliency methods, including D-RISE.
• To analyze the impact of different backbone architectures and datasets on explanation quality.

Main Methods:

• Comparison of gradient-based (e.g., guided backpropagation) and perturbation-based (e.g., D-RISE) saliency methods.
• Extensive experiments using diverse backbone architectures and datasets.
• Visual interpretation and numerical assessment of explanation methods.

Main Results:

• Both D-RISE and guided backpropagation produced localized explanations.
• D-RISE highlighted more semantically meaningful regions, leading to more human-understandable explanations.
• This study is the first to generate explanations focused on bounding box coordinate regression for object detection.

Conclusions:

• D-RISE provides superior, human-understandable explanations for autonomous driving object detection compared to other tested saliency methods.
• Explainable AI (XAI) is vital for building trust in deep learning models for safety-critical applications like autonomous driving.
• Further research into XAI for regression tasks in autonomous driving is warranted.