An Improved Regression-Based Method for Centerlines and Boundary Detection

In this work we present an extension of the regression-based method presented in [27] for centerline and boundary detection. As shown in [27], classification based approaches are inaccurate when applied to centerline detection of linear structures. The reason is that centerline points and locations immediately next to them are extremely difficult to distinguish, The solution proposed in [27] is to reformulate centerline detection as a regression problem. Instead of the binary function used by classification-based approach, the authors approximate a function designed to respond maximally along the centerlines and with smaller values on other pixels, depending on their distance to the centerlines. The idea of using regression instead of classification is general and can be applied to other problems with similar properties. Contour detection is one of these. Consider Fig. 2, where are shown the boundaries marked by 5 human subjects on a sample image. Even if the 5 humans most of the time agree on the detection of a boundary, the exact localization of this boundary is not precise and the annotations are distant by few pixels. This uncertainty can be caused by several factors, like for example lack of local information, low resolution, blurring and other image artifacts. Because of the ambiguity of pixels in a neighborhood of a boundary and their extremely similar aspect, it is not clear if pixels close to boundaries should be considered as positive or negative samples. As for centerline detection, training a classifier to distinguish boundary points from the rest may produce multiple responses on the boundaries and low localization accuracy. In this work we use an auto-context like approach to improve the performance of [27]. We test our method on the following two problems: 1. detecting road’s centerline in 2D aerial images; 2. contour detection in natural images. We