To exploit these graph structures, we are developing Graph Neural Networks (GNNs) based on the Transformer model. Another good way to visualize the behaviour of a function f(x,y) f ( x, y) is to sketch what are called its level curves. Thus, representing sketches with graphs offers a universal representation that can make use of both the sketch structure (like images) as well as temporal information (like stroke sequences). Often the reason you are interested in a surface in 3d is that it is the graph z f(x,y) z f ( x, y) of a function of two variables f(x,y). We assume that sketches are sets of curves and strokes, which are discretized by a set of points representing the graph nodes.Įach node encodes spatial, temporal and semantic information. GraphSketcher is described as OmniGraphSketcher helps you make elegant and precise graphs in seconds, whether you have specific data to visualize or you. We are working on a novel representation of free-hand sketches as sparsely-connected graphs.
temporal order: can we have the best of both worlds? Sketches as Graphs Recurrent Neural Networks (RNNs) stick out as a natural architecture for capturing this temporal nature of sketches. Sketches are usually an extremely sparse sequences of strokes which capture high-level abstractions and ideas. While CNNs are designed for static collections of pixels with dense colors and textures,
If we consider sketches as 2D images, we can throw them into off-the-shelf Convolutional Neural Networks (CNNs). Human beings have been creating free-hand sketches, i.e., drawings without precise instruments, since time immemorial.ĭue to the popularity of touchscreen interfaces, machine learning using sketches has emerged as an interesting problem with a myriad of applications: