Search for a command to run...
Object detection models based on deep learning achieve state-of-the-art accuracy in many applications, but their predictions are difficult to interpret. Existing explainable artificial intelligence (XAI) techniques for computer vision mostly focus on image classification, and their direct extension to object detection is non-trivial, as detectors must ex-plain both what was detected (class) and where it was detected (bounding box). In this paper, we revisit perturbation-based saliency methods for object detection, with a particular focus on D-RISE, a popular model-agnostic technique. We highlight practical limitations of D-RISE, particularly the absence of a detector-specific importance score aligned with detection performance. We then propose a novel detection-loss based explainability score that quantifies the importance of each image region through the deterioration of the detector’s loss under random masking. The proposed score jointly accounts for classification confidence and localisation quality, and it is normalised to be in [0, 1], providing an interpretable per-pixel importance measure. The proposed score is evaluated on three datasets covering industrial and general-object detection scenarios. In addition to qualitative visualization, we report object-level explainability statistics and deletion-based fidelity analysis to assess the alignment between highlighted regions and detection confidence. The re-sults indicate that the proposed ∆-based scoring method produces consistent and interpretable saliency maps that are directly linked to changes in detection loss. Overall, the proposed approach establishes a principled link between visual explanations and detection performance, and paves the way for robust, quantitative comparison of XAI methods for objectdetection.