DeepScores contains high-quality music score images divided into 300,000 written music scores containing music symbols of different shapes and sizes. The dataset contains nearly 100 million small objects in total, making it the largest public dataset. Ground truth annotations for object classification, object detection, and semantic segmentation are provided in the dataset.

DeepScores aims to advance research in the field of small object recognition by placing the object recognition problem in the context of scene understanding, while also raising related challenges in computer vision, especially optical music recognition research.