Automatic Image Generation Pipeline for Instance Segmentation of Deformable Linear Objects.

Area of Science:

• Robotics and Automation
• Computer Vision
• Machine Learning

Background:

• Automated handling and assembly of deformable linear objects (DLOs) like cables and hoses face challenges in robust detection.
• Limited availability of training data hinders the performance of deep learning models for DLO instance segmentation.

Purpose of the Study:

• To propose an automatic image generation pipeline for creating synthetic training data for DLO instance segmentation.
• To address the data scarcity issue in deep learning-based DLO detection for industrial applications.

Main Methods:

• Developed an automatic image generation pipeline allowing users to define boundary conditions for creating synthetic DLO training data.
• Modeled DLOs as rigid bodies with versatile deformations, identifying this as the most effective replication type.
• Defined reference scenarios for DLO arrangement to automatically generate diverse scenes within a simulation environment.

Main Results:

• Models trained on synthetically generated images demonstrated feasibility when validated on real-world images for DLO segmentation.
• The proposed data generation approach achieved results comparable to state-of-the-art methods.
• The pipeline showed significant advantages in reducing manual effort and enhancing transferability to new industrial use cases.

Conclusions:

• The automatic image generation pipeline effectively addresses the lack of training data for deformable linear object segmentation.
• This method offers a practical and efficient solution for improving the automation of tasks involving cables and hoses.
• The approach demonstrates strong potential for broad applicability across various industrial automation scenarios due to its transferability.