Automatic Instance Segmentation Labeling of Road Sign Image from Bounding Box Data

The rising popularity of artificial intelligence can be attributed to the amazing capabilities of supervised learning. This training method requires a vast amount of data to train the model. In computer vision, many datasets are publicly available for image classification and object detection tasks. However, the number of datasets available for image instance segmentation tasks could be improved, which shows how resource-intensive and complex the image labeling process is. Since many available image datasets are labeled for object detection, the preliminary task before instance segmentation, we propose a novel method for automatically labeling road marking signs based on bounding box-labeled data in this paper. Our method combines several image processing techniques to get the image labeled, for instance, segmentation. Tested on the Taiwan Road Marking Sign Dataset (TRMSD), our method can label more than 60% of the images. We are significantly cutting the time needed to label the images manually.