Linking Sketch Patches by Learning Synonymous Proximity for Graphic Sketch Representation

Graphic sketch representations are effective for representing sketches. Existing methods take the patches cropped from sketches as graph nodes, and construct edges based on sketch's drawing order or Euclidean distances canvas. However, of a may not be unique, while semantically related parts far away each other In this paper, we propose an order-invariant, semantics-aware method graphic representations. The linked according to their global semantics local geometric shapes, namely synonymous proximity, by computing cosine similarity between captured patch embeddings. Such constructed learnable adapt variation drawings, which enable message passing among patches. Aggregating messages convolutional networks plays role denoising, is beneficial produce robust embeddings accurate Furthermore, enforce clustering constraint over jointly with network learning. self-organized compact clusters, guided move towards assigned cluster centroids. It raises accuracy computed proximity. Experimental results show that our significantly improves performance both controllable synthesis healing.