Ceres: an Interactive Optical Music Recognition System

We present our Optical Music Recognition (OMR) system, Ceres, which embeds human intelligence into machine recognition. Different from previous OMR systems, which took human corrections as an extra post-processing step, our systems accepts user-supplied information during the recognition process. Due to the big challenge of OMR, all the existing systems suffer from different levels of errors. The inadequate recognition accuracy impeded the practical use of these systems, which always requests a large amount of human labor for error corrections. This proofreading process is time-consuming, laborious and sometimes error-prone. To alleviate the burden of proofreaders, we apply human-in-the-loop computation to encourage tight collaboration between human and computer, where each can contribute its best. In our system, user is only responsible for providing a small amount of information while the recognition and registration of symbols are left to the computer. We demonstrate our idea through an interactive OMR system, which shows the potential to considerably save the human efforts for OMR proofreading and be useful in practice.