Cable Instance Segmentation with Synthetic Data Generation

Abstract

We propose a bottom-up approach for the instance segmentation of cables (commonly referred in the literature as deformable linear objects). While the state of the art instance segmentation techniques propose a bounding box and perform foreground segmentation within each proposed bounding box, we adopt a bottom-up approach as cables can span a considerable part of the image or even the entire image, and therefore, cannot be well localized in a bounding box. In this paper, we show that several operations in the top-down instance segmentation approaches are only applicable for certain classes (i.e., compact objects) such as cars but they are a poor approximation for objects with highly overlapping bounding boxes such as cables. In particular, the non-maximum suppression and RoIPool/RoIAlign operations limit the generalizability of proposal-based instance segmentation methods to such datasets. Furthermore, we introduce a synthetic data generation technique that can also be applied to other popular public datasets such as COCO, Pascal VOC, and Cityscapes.