ASMOD: Adaptive Saliency Map on Object Detection

Although deep neural networks (DNNs) have achieved remarkable results in many fields, their interpretability is poor, and the decision-making process is not clear, which is often considered as a black-box model. In this paper, we propose a method Adaptive Saliency Map on Object Detection (ASMOD), multi-scale mask and suppression threshold are added to D-RISE black-box method for object detection task, which effectively improves the problem of fuzzy decision region and poor interpretability when generating saliency map (estimating the prominence of each pixel model prediction). Meanwhile, we applied the improved method to the four target detectors with different structures for explanation, analysis and comparison, and gave suggestions for improvement. We use two metrics: Automatic Deletion / Insertion metric and Pointing game metric based on manual annotation to validate our improved performance, the results show that our method can better adapt to multi-scale targets and reduce irrelevant parts, thus enhancing the interpretability of the model and the credibility of the saliency map.