Auto-generated Wires Dataset for Semantic Segmentation with Domain-Independence

Abstract

In this work, we present a procedure to automatically generate an high-quality training dataset of cable-like objects for semantic segmentation. The proposed method is explained in detail using the recognition of electric wires as a use case. These particular objects are commonly used in an extremely wide set of industrial applications, since they are of information and communication infrastructures, they are used in construction, industrial manufacturing and power distribution. The proposed approach uses an image of the target object placed in front of a monochromatic background. By employing the chroma-key technique, we can easily obtain the training masks of the target object and replace the background to produce a domain-independent dataset. How to reduce the reality gap is also investigated in this work by correctly choosing the backgrounds, augmenting the foreground images exploiting masks. The produced dataset is experimentally validated by training two algorithms and testing them on a real image set. Moreover, they are compared to a baseline algorithm specifically designed to recognise deformable linear objects.