Generating Synthetic Image Segmentation Dataset Using Realistic City Backgrounds
İrem İşlek, N. Deniz Aksaylı, Onur Güngör, Çağla Çığ Karaman · 2022 30th Signal Processing and Communications Applications Conference (SIU) · 2022
Creating a dataset for the image segmentation problem is a time-consuming process. For this reason, generating synthetic datasets that are quite close to reality has great importance. In this study, for the problem of segmentation of various objects encountered in the city, a method for synthetically generating a set of images that contain objects in an city background employing images that do not have city backgrounds is proposed. This method creates a synthetic segmentation dataset of the desired size with realistic city backgrounds, using city videos from video-sharing platforms and a segmentation dataset related to the object to be segmented. The proposed approach has been tested for a pedestrian segmentation problem and the detailed results have been compared with the results of a real dataset. Results show that using the synthetic dataset increases the AP value by 8 percentage points and the AR value by 4 percentage points compared to not using it.