Limelight allows you to run object detection pipelines when using a Google Coral USB Accelerator. You can either upload a pretrained game piece detection model (which can be found here) or you can train a tensorflow lite model on your own dataset.
First, create your dataset. You can take your own images, or you can find a datatset on Roboflow Universe. Next, download the dataset in Tensorflow TFRecord format. Now that you have your dataset, you can fine-tune a pre-trained limelight model for your dataset. Follow the instructions in this notebook to do so. Now you can upload your model to your limelight to start running object detection.