Details


We created PTAD with three main aims:

  • to serve as a synthetic benchmark for texture analysis models and algorithms,
  • to be sufficiently complex to allow for some generalisation to real-world images, and
  • for the size of the images, number of classes, and amount of textures to be easily tunable.

The images contained in the dataset are created using a combination of simplex noise, connected components, the convex hull algorithm, and the Describle Textures Dataset (DTD); among some other approaches to noise generation, blurring, and smoothing.


The following table shows the test results from 5 different FCN architectures trained on PTAD using a non-semantic loss function called Hungarian Cross Entropy:

Architecture Adjusted Rand Index Adjusted Mutual Information Hungarian Cross Entropy
U-Net 0.42 0.46 0.98
Dilation 2 0.41 0.42 0.93
Dilation 3 0.60 0.58 0.69
B-Net 0.85 0.84 0.23
TSC-Net 0.91 0.89 0.15

GitHub


The Github Repo contains some things you may find useful:


Download


Below, you can download the training, validation, and testing sets we used; so that you can benchmark your architecture against ours. The training, validation, and testing sets contain 80000, 20000, and 5000 samples, respectively; there is a maximum of 5 clusters in each image.