AN INFINITELY SCALABLE DATASET OF SINGLE-POLYGON GRAYSCALE IMAGES AS A FAST TEST PLATFORM FOR SEMANTIC IMAGE SEGMENTATION

Abstract

Background. Every new semantic image segmentation task requires fine-tuning the segmentation network architecture that is very hard to perform on images of high resolution, which may contain many categories and involve huge computational resources. So, the question is whether it is possible to test segmentation network architectures much faster in order to find optimal solutions that could be imparted to real-world semantic image segmentation tasks.Objective. The goal of the article is to design an infinitely scalable dataset, which could serve as a test platform for semantic image segmentation. The dataset will contain any number of entries of any size required for testing.Methods. A new artificial dataset is designed for semantic image segmentation. The dataset is of grayscale images with the white background. A polygonal object is randomly placed on the background. The polygon edges are black, whereas the polygon body is transparent. Thus, a dataset image is a set of edges of a convex polygon on the white background. The polygon edge is one pixel thick but the transition between the white background and the polygon black edges includes gray pixels in the vicinity of one-pixel edges. Such a noise is an aftermath of the image file format conversion process. The number of edges of the polygon is randomly generated for every next image. The polygon size and position of its center of mass with respect to image margins are randomized as well.Results. A toy dataset of any volume and image size from scratch can be generated. Besides, the dataset generator automatically labels pixels to classes “background” and “polygon”. The dataset does not need augmentation. Eventually, the dataset is infinitely scalable, and it will serve as a fast test platform for segmentation network architectures.Conclusions. The considered examples of using the polygonal dataset confirm its appropriateness and capability of networks trained on it to successfully segment stacks of objects. Additionally, a criterion of early stopping is revealed based on empty image segmentation.