Lin Yancong ; Pintea Silvia L. ; van Gemert Jan C.

Abstract

Classical work on line segment detection is knowledge-based; it usescarefully designed geometric priors using either image gradients, pixelgroupings, or Hough transform variants. Instead, current deep learning methodsdo away with all prior knowledge and replace priors by training deep networkson large manually annotated datasets. Here, we reduce the dependency on labeleddata by building on the classic knowledge-based priors while using deepnetworks to learn features. We add line priors through a trainable Houghtransform block into a deep network. Hough transform provides the priorknowledge about global line parameterizations, while the convolutional layerscan learn the local gradient-like line features. On the Wireframe(ShanghaiTech) and York Urban datasets we show that adding prior knowledgeimproves data efficiency as line priors no longer need to be learned from data.Keywords: Hough transform; global line prior, line segment detection.