Beyond Fixed Thresholds and Domain-Specific Benchmarks for Explainable Multi-Task Classification in Autonomous Vehicles

Deep learning models power scene understanding in autonomous vehicles, but their black-box nature raises transparency and safety concerns. A new paper from Maryam Sadat Hosseini Azad and Shahriar Baradaran Shokouhi tackles this with a comprehensive confidence threshold sensitivity analysis for multi-task explainable classification. Instead of relying on fixed thresholds, they introduce an adaptive approach that dynamically selects optimal decision boundaries for each driving task. Their experiments show that this method consistently improves F1-scores across multiple tasks, proving that fixed thresholds are suboptimal in multi-task scenarios. This makes autonomous perception systems both more accurate and more interpretable, a critical step for human trust and regulatory approval.

Alongside the methodological advance, the authors present IUST-XAI-AD, a new dataset of 958 images with human annotations for driving decisions and their underlying reasoning. Unlike existing benchmarks, this dataset captures diverse driving contexts and cross-cultural behavior patterns, providing a more challenging and realistic test environment. The combination of adaptive thresholding and this dataset reveals important insights about how driving behaviors vary across cultures, enabling the development of safer, more reliable autonomous systems that can be deployed globally. This work bridges the gap between technical performance and explainability in autonomous driving, offering both better algorithms and better evaluation tools.