Dataset and Performance Metrics towards Semantic Segmentation

Abstract

Interest and ideas on semantic segmentation move on the increasing trend in the area of autonomous driving. This meets the rise in the deep learning approach. The first step in the training of a segmentation model is the dataset preparation. For this, RGB images and its corresponding segmentation images are required such that, the size of these remain the same. Each class in the image is assigned with a unique ID. The pixel value in the segmentation image denotes the class ID of the corresponding pixel. Moreover, as jpg format of the image is lossy, bmp or png formats are usually preferred. The success of the model is measured using metrics, which helps in grading the model. This paper deals with the examination of the widely used datasets in the field of semantic segmentation. The mIoU metric of the datasets on various models have been comparative studied at the end of the analysis.