Robustness Analysis of YOLOv9 Object Detection Under Dynamic Illumination Variance for Museum Artifacts

Abstract

Object detection technology has been widely applied in smart tourism and digital museum systems. However, variations in illumination conditions pose a significant challenge, affecting the visibility of objects and the object localization performance. This study aims to evaluate the robustness of the YOLOv9 model in detecting museum objects under different illumination conditions. To evaluate the model performance, 5-Fold Cross-Validation was used on a multi-class dataset of eleven classes of cultural heritage artifacts from Museum Lampung. To test the effects of illumination variation, testing was done under three different lighting conditions (normal lighting, 50 percent dimmed lighting, and 50 percent brightened lighting). The results show that YOLOv9 can maintain stable detection performance in such conditions, with a mean Average Precision at 50 percent (mAP@50) of 0.991 and a mean mAP at 50 to 95 percent (mAP@50-90) of 0.846. In particular, under normal lighting conditions, the model achieved a mAP@50 of 0.995 and a mAP@50-90 of 0.946. For under 50 percent dimmed and brightened lighting, mAP@50-90 dropped to 0.928 and 0.933, respectively. These results suggest that illumination variation affects the localization accuracy more under stricter Intersection over Union thresholds than the general object detection performance. Overall, the results of this research work demonstrate that YOLOv9 can maintain stable object detection performance in a museum environment despite variations in illumination conditions.