Rotating Features for Object Discovery

This repository contains the 4Shapes datasets from the paper "Rotating Features for Object Discovery" (link) by Sindy Löwe, Phillip Lippe, Francesco Locatello and Max Welling.

The 4Shapes dataset comprises grayscale images of dimensions 32 x 32, each containing four distinct white shapes (square, up/downward facing triangle, circle) on a black background. The dataset consists of 50,000 images in the train set, and 10,000 images for the validation and test sets, respectively. All pixel values fall within the range [0,1].

The 4Shapes RGB(-D) datasets follow the same general setup, but randomly samples the color of each shape. To achieve this, we create sets of potential colors with varying sizes. Each set is generated by uniformly sampling an offset value within the range [0,1], and subsequently producing different colors by evenly dividing the hue space, starting from this offset value. The saturation and value are set to one for all colors, and the resulting HSV color representations are converted to RGB. To create the RGB-D variant, we incorporate a depth channel to each image and assign a unique depth value within the range [0,1] to every object, maintaining equal distances between them.

For more details on how to use these datasets, see our GitHub repository.