ELSM: Evidence-Based Line Segment Merging

Abstract

Abstract Existing line segment detectors break perceptually contiguous linear structures into multiple line segments. This can be offset by re-merging the segments, but existing merging algorithms over-merge and produce globally incorrect segments. Geometric cues are necessary but not sufficient for deciding whether to merge two segments or not. By restricting the result of any merging decision to have underlying image support, we reduce over-merging and globally incorrect segments. We propose a novel measure for evaluating merged segments based on line segment Hausdorff distance. On images from YorkUrbanDB, we show that our algorithm improves both qualitative and quantitative results obtained from four existing line segment detection methods and is better than two existing line segment merging methods. Our method does not suffer from inconsistent results produced by four recent deep learning-based models. The method is easily customisable to work for line drawings such as hand-drawn maps to obtain vectorized representations.