v1.0.2

• Determine how many misclassified positions are illegal according to the rules of chess. The results are obtained by running python3 chesscog.report.analyze_misclassified_positions --results data://recognition --dataset test and python3 chesscog.report.analyze_misclassified_positions --results data://transfer_learning/recognition --dataset test):
  • In the test set (342 samples total), 21 positions were misclassified, of which 1 was illegal according to the rules of chess.
  • In the test set of the transfer learning dataset (27 samples total), 3 positions were misclassified, of which all 3 were illegal according to the rules of chess.
• Display a warning in the inference scripts (chesscog.recognition.recognition and chesscog.transfer_learning.recognition) if the predicted position is illegal according to the rules of chess.

Note: FEN positions are validated using the chess library.

v1.0.1

• Move data storage from Google Drive to OSF; the download script now downloads and unpacks the data from OSF.

v1.0.0

First stable version!

• Implement transfer learning algorithm to adapt the recognition system to an unseen chess set based on only two images of the starting position.
• Evaluate transfer learning approach.
• More scripts to evaluate the results for the report.
• Write a lot more documentation.
• Auto-generate documentation using Sphinx. Docs are available at https://georgw777.github.io/chesscog.
• More tests.

v0.2.7

• compute more evaluation statistics
• improve performance of chessboard corner detection
• implement new data augmentations (shear, scale, translate)
• implement transfer learning capabilities (achieve >99.9% per-square accuracy for both classifiers on unseen images, using only two training images of the starting chess position from both players' perspectives)

v0.2.6

• run Canny edge detection over edge gradients for border refinement
• end-to-end evaluation

v0.2.5

• more generous thresholds for corner detection

v0.2.4

• better RANSAC parameters

v0.2.3

• eliminate far-away lines in corner detection

v0.2.2

• extract configuration management to the new recap package
• implement border refinement via edge gradients
• implement image resizing
• implement diagonal line rejection

v0.2.1

• implement CI with chesscog-app
• automated deployment to the web