Figure 1: We parse a 2D vector graphics image (root node of the tree on the left) into a perceptually organized hierarchy of graphical elements, and use it to power selection tools. We collect a dataset of such hierarchies and train a neural network that learns to group by recursively merging elements. One of our selection tools (top right) works by selecting an initial path (the swimmer's face in this example) and using the plus button to traverse to ancestors of the path. The expanded selection is highlighted in light blue. Competing state-of-the-art approaches, Fisher et. al. [7] (inset bottom left) and Suggero [26] (inset bottom right), fail to grow the selection correctly, grouping head and ball in one case and omitting all outlines in the other. This is because handwritten heuristics alone are not robust indicators of grouping affinity: our data-driven method learns to avoid such errors.