Generator of a Toy Dataset of Multi-Polygon Monochrome Images for Rapidly Testing and Prototyping Semantic Image Segmentation Networks

Abstract

In the paper, the problem of building semantic image segmentation networks in a more efficient way is considered. Building a network capable of successfully segmenting real-world images does not require a real semantic image segmentation task. At this stage, called prototyping, a toy dataset can be used. Such a dataset can be artificial and thus may not need augmentation for training. Besides, its entries are images of much smaller size, which allows training and testing the network a way faster. Objects to be segmented are one or few convex polygons in one image. Thus, a toy dataset generator is created whose complexity is regulated by the number of edges in a polygon, the maximal number of polygons in one image, the set of scale factors, and the set of probabilities determining how many polygons in a current image are generated. The dataset capacity and image size are concurrently adjustable, although they are much less influential.

Image semantic segmentation method based on improved ERFNet model

“…But the bottleneck structure is often affected by degradation problems. Considering that each convolutional layer can be decomposed by a combination of 1D filters, the resulting low-dimensional decomposition layer has a simple structure and can reduce computational cost [34,35]. For this reason, a 1D filter is used to redesign the bottleneck module, and the improved residual layer structure is called a weak bottleneck module, as shown in Figure 3.…”