Holistically-Attracted Wireframe Parsing

This paper presents a fast and parsimonious parsing method to accurately androbustly detect a vectorized wireframe in an input image with a single forwardpass. The proposed method is end-to-end trainable, consisting of threecomponents: (i) line segment and junction proposal generation, (ii) linesegment and junction matching, and (iii) line segment and junctionverification. For computing line segment proposals, a novel exact dualrepresentation is proposed which exploits a parsimonious geometricreparameterization for line segments and forms a holistic 4-dimensionalattraction field map for an input image. Junctions can be treated as the"basins" in the attraction field. The proposed method is thus calledHolistically-Attracted Wireframe Parser (HAWP). In experiments, the proposedmethod is tested on two benchmarks, the Wireframe dataset, and the YorkUrbandataset. On both benchmarks, it obtains state-of-the-art performance in termsof accuracy and efficiency. For example, on the Wireframe dataset, compared tothe previous state-of-the-art method L-CNN, it improves the challenging meanstructural average precision (msAP) by a large margin ($2.8\%$ absoluteimprovements) and achieves 29.5 FPS on single GPU ($89\%$ relativeimprovement). A systematic ablation study is performed to further justify theproposed method.