Explainable Multi-Camera 3D Object Detection with Transformer-Based Saliency Maps

Explainability is critical for autonomous driving safety: The paper addresses a key gap in deploying Vision Transformers for 3D object detection in autonomous vehicles, where understanding why the AI makes decisions is essential for regulatory approval, developer accountability, and user trust.

Attention-based explanations are more practical than existing methods: The proposed saliency map approach based on transformer attention is more computationally efficient than gradient-based explainability methods, making it more feasible for real-world deployment in safety-critical systems.

Multi-layer analysis improves explanation quality: The research demonstrates that aggregating attention information across different layers of the transformer produces better explanations than single-layer approaches, suggesting that AI transparency requires looking at the full decision-making process rather than isolated components.

Explainability methods can be rigorously validated: The paper uses perturbation tests and quantitative metrics on the nuScenes dataset to evaluate explanation quality, showing that explainability itself can be measured and compared—not just assumed to work—which is important for responsible AI deployment.

Transparency can increase stakeholder confidence: By making Vision Transformer reasoning visible through saliency maps, this work supports the broader goal of building trust among authorities, developers, and users in AI systems used for high-stakes applications like autonomous driving.