“ Multiscale Centerline Detection by Learning a Scale - Space Distance Transform ” - Centerline Detection

In this appendix we present the results obtained by applying our method to the problem of boundary detection in natural images. We show that we can outperform state-of-the-art techniques also on this task. Many analogies connect the centerline detection to the contour detection problem. First, in both cases we want to extract one dimensional curves form images. Second, in both situation the exact localization of a boundary or of a centerline is not well defined and uncertainty exists also in the human annotated ground truth. For example in Fig. 1 we see an image annotated by 5 human subjects. Even if the subjects agree in the detection of a boundary, the localization of this boundary differs from human to human by few pixels.