CAESAR++: Uncertainty-Driven Contextual Reasoning for Trustworthy and Explainable Road Object Detection

Anh-Thu Mai, Marina Nicolas, Patricia Ladret, Alice Caplier

Keywords: Explainable artificial intelligence, road object detection, conformal prediction, contextual reasoning

TL;DR: Look around when in doubt for safer, smarter, and more explainable road-object detection

Abstract:

This paper introduces CAESAR++, a framework for road object detection that combines conformal prediction, adaptive contextual reasoning, and dual-colour saliency maps. CAESAR++ first calibrates classification and localisation uncertainty using a two-step conformal
procedure, then enlarges the context window around each detection in proportion to its uncertainty, and finally produces object-wise explanations that disentangle bottom up sensory evidence from top-down contextual cues. Experiments indicate consistent improvements in detection accuracy, uncertainty calibration, and explanation stability without retraining the base models.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 7