Graph Kernel Matching

Multi-level Correspondence via Graph Kernels for Editing Vector Graphics Designs

To create graphic designs such as infographics, UI mockups, or explanatory diagrams, designers often need to apply consistent edits across similar groups of elements which is a tedious task to perform manually. One solution is to explicitly specify the structure of the design upfront and leverage it to transfer edits across elements that share the predefined structure. However, defining such a structure requires a lot of forethought which conflicts with the iterative work- flow of designers. We propose a different approach where designers select an arbitrary set of source elements, apply the desired edits, and automatically transfer the edits to similarly structured target elements. To this end, we present a graph kernel-based algorithm that retroactively infers the shared structure and correspondence between source and target elements. Our method does not require any explicit annotation and can be applied to any existing design regardless of how it was created. It is flexible enough to handle differences in structure and appearance between source and target graphics, such as cardinality, color, size, and arrangement. It also generalizes to different types of edits such as style transfer or applying animation effects. We evaluate our algorithm on a range of real-world designs and demonstrate how our approach can facilitate various editing scenarios.