Oblique Decision Trees as an Image Model for Cubist Image Restyling

We propose the use of oblique decision trees and forests, originally a supervised machine learning approach, as a model of a single image. This results in a hierarchical partition into constant-color polygons ("poxels") that best approximates the source image. The tree or forest can be trained using a recent algorithm that applies alternating optimization over the nodes of a tree and the trees of a forest. In this paper, we use this for aesthetic effects, such as restyling an image to look like a cubist painting or a stained-glass window. We illustrate this with multiple examples and compare it with the results achieved by Generative AI and Neural Style Transfer.