This project turns before/after camera frames of a physical chessboard into move predictions by recovering board pose, rectifying the image, computing a diff, and scoring square-to-square moves.

What it demonstrates

• computer-vision tracking from ordinary hardware instead of instrumented boards
• board-pose recovery and perspective correction before model inference
• a practical bridge from synthetic data generation to end-user tooling

Highlights

• Blender-based synthetic dataset generation for varied boards, pieces, and lighting
• PyTorch move prediction models, including the newer diff-based approach
• end-user surfaces for both phone-camera capture and move recording workflows
• write-up and repository aligned to the current v0.2.0 release

Links

• Read the write-up
• View the code on GitHub