Graph-based geometric structure line parsing

Line drawings by artists capture the organization, relationships, and semantics of observable objects. To endow machines with similar capacities and improve the storage and processing of such drawings, we investigate uniform representation and learning of hetero-dimensional line drawings. For this, nonlinear curves are first approximated as paths, resulting in node detection and link prediction (LP). An end-to-end trainable neural network is then developed for parsing structural-semantic sketches from images based on such representations. In particular, node regression and feature extraction are performed using convolutional neural networks(CNNs), and the prediction of connections is achieved using a graph-based encoder-decoder architecture. We provide a topology-guided graph initialization strategy, which increases the accuracy of LP but also greatly minimizes the self-overlapping among predicted paths. We also develop a semi-automatic toolkit for annotation generation, as well as a dataset for understanding structural sketches-the first large parametric dataset consistent with human visual perception. Through comparative experiments and ablation studies, we demonstrate the advantages and effectiveness of our methods.