Zhou Yichao ; Qi Haozhi ; Ma Yi

Abstract

We present a conceptually simple yet effective algorithm to detect wireframesin a given image. Compared to the previous methods which first predict anintermediate heat map and then extract straight lines with heuristicalgorithms, our method is end-to-end trainable and can directly output avectorized wireframe that contains semantically meaningful and geometricallysalient junctions and lines. To better understand the quality of the outputs,we propose a new metric for wireframe evaluation that penalizes overlapped linesegments and incorrect line connectivities. We conduct extensive experimentsand show that our method significantly outperforms the previousstate-of-the-art wireframe and line extraction algorithms. We hope our simpleapproach can be served as a baseline for future wireframe parsing studies. Codehas been made publicly available at https://github.com/zhou13/lcnn.