Generation of realistic synthetic cable images to train Deep Learning segmentation models

One of the main contemporary challenges in robotics is the automation of tasks that involve handling Deformable Linear Objects (DLOs), such as cables or ropes. Due to their changing shape, constant perception is required, which makes computer vision, and in particular, image segmentation, fundamental. Even though image segmentation is a very studied problem, which has been addressed with high accuracy by many different Deep Learning models, they need to be properly trained to segment the objects of interest (DLOs in this case). To do this, it is necessary to have a large and diverse dataset. However, such a dataset hasn’t been created for cable images, and doing it manually would be complicated and extremely time-consuming. This paper addresses this issue with a novel methodology, which automatically generates synthetic cable datasets to train image segmentation models. This methodology utilizes Blender to create photo-realistic scenes and a Python pipeline to interact with them. To ensure the diversity of the dataset, before the generation of each image, the pipeline performs random variations on the elements of the scene. After this, the scene is animated, dropping the cables from a certain height and letting them deform naturally after landing, which results in realistic arrangements of the cables. The effectiveness of the methodology was demonstrated by training six popular segmentation models with synthetic datasets and using them to segment real cable images with great results (IoU over 70% and Dice coefficient over 80% for all the models).