Shapetastic OOD

This dataset was created with the Shapetastic generator and has multiple partitions as follows. First, a partition with 50k grayscale images with a white background and black shapes drawn from each of the five available shapes, where the outline varies in width and is always gray. Second, a partition with 50k images similar to the previous, except that colors of both background and shapes are randomly selected from a color palette comprising 148 colors.Again, the outline varies in width, but this time always in the interpolated color between the background and the color of the shape.With this partition, we can for instance assess the impact of the presence of colors compared to the grayscale images. Third, a partition with 60k grayscale images with a white background and black elongated rectangles, ellipses, and parallelograms where the outline varies in width and is always gray.Here, the focus lies on more fine-grained rotations every five degrees between 0 and 180 degrees.We needed to constrain some parameters to make sure the shapes were oriented according to the ground truth, which is for instance not possible if we would include polygons and stars.This partition allows us to unveil a bias in the methods that is rooted in the predominant edge orientation of the image. Each of these partitions have a resolution of 200x200 pixels. Last, two partitions that each contain 50k images of 5 different shapes without interpolations at the outlines. One dataset was generated directly with 32x32 pixels while the other was generated with 64x64 pixels. These partitions allow us to investigate any side effects of resizing to 32x32 pixels.