Early Warning of Performance Degradation in Object Detection Models via Explainable AI

초록

In real-world deployment environments, visual degradations such as fog, motion blur, and low/high illumination frequently occur, potentially leading to significant performance degradation in object detection models. However, most existing studies primarily focus on post-hoc performance evaluation metrics, such as mean Average Precision (mAP), which only capture performance drops after they occur. In this work, we investigate whether explainable AI-based internal indicators can provide early warning signals of performance degradation in object detection models. To analyze the behavior of an object detection model before degradation becomes observable in mAP , we examine internal indicators at three levels - Input, Decision, and Concept - under controlled visual corruptions. We compare how these indicators respond as corruption severity increases and whether they change earlier than mAP. The results show that some IDC indicators, particularly at the Input and Concept levels, change before degradation becomes clearly observable in mAP. © 2026 IEEE.

키워드