This week was midterm-heavy, so most of my project work took place later in the week. After exams, May and I focused on completing all soldering for the LED connections and repairing previously broken ones. While we secured the joints with electrical tape, several connections still came loose during testing, so we plan to explore more stable soldering or wire management methods next week.
In parallel, I continued developing the dice detection algorithm. I continue experimenting with multiple approaches, such as background subtraction, bounding box isolation, and DBSCAN clustering, to improve pip recognition accuracy. I successfully integrated the algorithm with the external Etron camera, enabling proper real-time testing. However, differentiating the dice from the surrounding background remains a challenge. Background removal often makes the dice less visible against the plate, reducing detection accuracy below our stated use-case requirement. A sample image of my current progress is below. I also placed orders for extra parts for our new design and second board.
After researching more effective techniques, I realized that pretrained models may not generalize well for our setup. Instead, I plan to train a custom YOLO model using images captured from our specific camera and lighting conditions. I will begin collecting training data once one board is fully assembled, which should be within the next two weeks. With midterms over, I also expect to help complete the second board next week and further refine the dice detection accuracy.
