Stochastic Context Models Designed for the Evaluation of Deep Generative Models of Images
Year: 2023
Type: dataset
Abstract: Numerous realizations from three different stochastic context models (SCMs) are given as 256x256-pixel, 8-bit grayscale image files in PNG format. These images can be used as training data for any deep generative model (DGM). Each SCM was designed to exhibit specific contextual features that can be measured in a post-hoc analysis of images generated by the trained DGM. The goal is to quantify the rates at which a given DGM architecture fails to reproduce the prescribed features. More details about the SOMs and their use can be found in the manuscript pre-print: A Method for Evaluating Deep Generative Models of Images via Assessing the Reproduction of High-order Spatial Context. SCM: Flags Eight different classes of "flags" are given. Each realization can be thought of as a grid of 16x16 pixel squares. Each class has precisely 80 squares set as bright foreground and 176 squares set as dark background. There are 24 squares that are never foreground in any class. The grayscale value of each pixel within the foreground is a random variate chosen from a scaled Beta distribution with parameters: a=4, b=2, location=96, scale=152. Analogously, each pixel within the background is a random variate chosen from a scaled Beta distribution with parameters: a=2, b=4, location=8, scale=192. All grayscale values throughout each realization are randomly arranged and were rounded to an integer value prior to placement. SCM: Voronoi Four different classes of Voronoi tilings are given. Classes are distinguished by the total number of tiles (N) which ranges from 16 to 64, inclusively, in increments of 16. For each realization, N 8-bit grayscale shades are chosen from a uniform distribution of values running from 12 to 251. Each tile is given precisely one of the chosen shades. Crucially, the grayscale value of the tiles increases monotonically with tile area. In other words, the shade and area are perfectly rank-correlated in every realization. SCM: Alphabet An eight-letter alphabet comprises the letters H, K, L, V, W, X, Y, Z which were chosen for visual distinctiveness. A letter is equally sized and centered within a 32x32 pixel tile. Each 256x256 realization is made as a raster of 64 letters into an 8x8 grid where each letter is one of the 32x32-pixel tiles. For every realization, the frequencies of the letters are identical: 24-H, 2-K, 16-L, 1-V, 1-W, 8-X, 8-Y, and 4-Z. The order of letters is random wherever a higher-order letter-arrangement rule is not prescribed. These rules are: 1) The letter pair X-Y is always horizontally adjacent (within the same row) in the 8x8 grid. 2) The ordered letter pairs Z-K, Z-V, Z-W are always vertically adjacent (within the same column) in the 8x8 grid. 3) The ordered letter pairs X-Y, Z-K, Z-V, Z-W have exact prevalences of 8, 2, 1, 1 respectively within each realization.
Source: Harvard Dataverse
DOI: 10.7910/dvn/hhf4af
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4398724821