We propose a new approach to semi-automated delineation of curvilinear structures in a wide range of imaging modalities. Earlier approaches lack robustness to imaging noise, do not provide radius estimates for the structures and operate only on single channel images. In contrast, ours makes use of the color information, when available, and generates accurate centreline location and radius estimates with minimal supervision. We demonstrate this on a wide range of datasets ranging from a 2D dataset of aerial images to 3D micrographs of neurites.