The model doesn't work
Hi, I was trying with some chessboards and the model was not returning proper FEN values, is it normal ?
I was also wondering, what are your training data ? Are you interested on some collaboration ?
Hi Beegbrain!
Thanks for trying out the chess FEN generation model and for the feedback! You're absolutely right about the accuracy issues - let me explain the current limitations.
Model Specificity & Limitations:
- The model was trained on a very specific dataset of 2D chessboard images (not 3D/angled boards)
- Board segmentation works reliably only on similar board styles to the training data
- YOLO piece detection models were trained on one specific piece set design
- Post-processing logic is tuned for this particular setup
Project Context:
This was just a side project I built to solve my own specific problems with chess position recognition. The models are quite narrow in scope and may not fit well for general use cases.
Technical Pipeline:
- Board segmentation for corner detection and perspective correction
- Custom YOLO models trained on my piece dataset
- Specialized post-processing logic for FEN completion
- ONNX optimization for my specific hardware setup
For Your Use Case:
If you're working with different board styles or piece sets, you'd likely need to retrain the models on your specific data. The architecture could work, but the trained weights are quite specialized.
Collaboration:
I'm definitely interested in collaboration on any topics! Whether it's chess-related projects, computer vision, AI model development, or completely new ideas - I'd be happy to discuss and work together on interesting projects.
Full Disclosure:
The entire model development, code implementation, and even this reply were created with AI assistance.
Thanks for the interest! 🎯
Nice, thank you for the detailed answer. I would like to build a proper pipeline for chess FEN generation that can be robust to new chessboards and pieces. Would you be interested for a training on larger scale dataset to improve the quality of the model inference ? There exist some board dataset online and I wanted to gather some of them and eventually a synthetic pipeline to even augment more the data, what do you think about this ?
is there a plateform on which you want to communicate ? (discord, mail, other ?)