A dataset containing the images and labels for the Chesseract data used in the CVPR NAS workshop Unseen-data challenge under the codename "Chester"
The Chesseract dataset is constructed from publicly accessible Chess games played by Grandmasters: Bobby Fischer, Garry Kasparov, Magnus Carlsen, Viswanathan Anand, Hikaru Nakamura, Anatoly Karpov, Fabiano Caruana, and Mikhail Tal.
We extracted the final 15% of board states where the nth board state refers to the position after N moves.
Chesseract uses 12 channels to encode the information rather than the traditional 2 or 3 for grayscale and RGB images respectively. Each channel represents a singular piece type and colour combination. In essence, this has allowed us to represent the pieces and their position on the board in a one-hot encoded format.
A rendering of one of the board states has been included in this archive to demonstrate how the machine perceives the dataset. A 2D traditional of the same board state is also included. Please note in the 3D rendering the white pieces have been rendered as gray to be more visible.
The data is in a channels-first format with a shape of (n, 12, 8, 8) where n is the number of samples in the corresponding set (49,998 for training, 10,001 for validation, and 10,001 for testing).
There are three classes in the dataset, with labels23,333 examples of each distributed as evenly as possible between the three subsets.
The three classes and corresponding numerical labels are as follows:
White wins: 0,
Black ins: 2