CL-DLO: Direct Centerline Prediction for Instance Segmentation of Deformable Linear Objects in Confined Environments

Abstract

Instance segmentation of cables in telecommunication ducts supports robotic inspection and maintenance by enabling individual cables to be identified and targeted. However, visual similarities between cables, close spacing, occlusions, and strong perspective distortion limit the application of existing methods, which often lead to the merging or incorrect identification of cable trajectories. To address these shortcomings, we propose CL-DLO (CenterLine Prediction of Deformable Linear Objects), a pipeline that directly predicts DLO centerlines from RGB images to preserve separation between nearby cables. A DINOv3-based architecture is employed to predict width-adaptive cable skeleton masks, and geometric endpoint matching reconnects fragmented paths into individual cable instances. For inspection videos, transverse observations are stacked into a spatio-temporal image for cable reconstruction. CL-DLO achieves an instance-level F1 score of 94.23%, compared to 90.11% for the strongest baseline evaluated on generic public benchmarks, and delivers up to a 2.2× improvement on real-world duct inspection images (70.7% versus 32.15%). For video-based inspection, CL-DLO reconstructs cable instances in both real and synthetic duct sequences, achieving sequence-level scores of 79.4% in a challenging synthetic case containing up to four partially occluding DLOs. These results present new benchmarks for underrepresented DLO perception settings, highlighting the potential of CL-DLO for application in real robotic inspection scenarios.

David Froelicher

Research Manager

My research interests include applied cryptography, AI, distributed systems and genomic privacy.