The MSRA-TD500 image dataset is used to evaluate the text detection algorithm benchmark. It contains 500 natural images of various categories and languages, including 300 training images and 200 test images.

The dataset is captured from indoor and outdoor scenes using a pocket camera. Indoor images are mainly signs, door panels, and warning signs, while outdoor images are mainly guide plates and billboards under complex backgrounds. The image resolution ranges from 1296×864 to 1920×1280.

The images contain Chinese and English, and each picture is fully annotated in lines, with difficult-to-identify ones having difficulty annotating them.

The MSRA-TD500 dataset was released by Huazhong University of Science and Technology at CVPR in 2012. The main publishers are C. Yao, X. Bai, W. Liu, Y. Ma and Z. Tu.