Optical Character Recognition (OCR) approaches have been widely advanced in recent years thanks to the resurgence of the
deep learning community. The state-of-the-art models are mainly trained on the datasets consisting of the constrained scenes
that involve considerable processing by human annotators. Detecting and recognizing text from the real-world images remains a technical challenge. In this paper, we introduce a large-scale OCR dataset Uber-Text,
which contains street-level images collected from car mounted sensors and truths annotated by a team of image analysts. The characteristics of the dataset include
(1) streetside images with their text region polygons and the corresponding transcriptions, (2) 9 categories indicating the business name text, street name text and street number text, etc,
(3) a set containing over 110k images,
(4) 4.84 text instances per image on average.
We show the challenge of the task and the dataset via evaluating two recently proposed object detection methods, which proves the significance of the dataset and motivates the future work in this field of study.
In addition, we propose an end-to-end text sequence recognition method which is lexicon-free and does not require character-level pretrained stage.
The dataset is distributed as Uber-Text.zip (197 Gb), and is split in training, validation and testing subsets. Furthermore, each subset is subdivided in 1K and 4K images. The dataset is distributed under Attribution-ShareAlike Version 4.0 International license .
Disclaimer: We have performed best-effort privacy blurring for faces and license plates in the imagery.
When using this dataset in your research, we will be happy if you cite us.
@inproceedings{UberText, author= {Ying Zhang and Lionel Gueguen and Ilya Zharkov and Peter Zhang and Keith Seifert and Ben Kadlec}, title = {Uber-Text: A Large-Scale Dataset for Optical Character Recognition from Street-Level Imagery}, booktitle = {SUNw: Scene Understanding Workshop - CVPR 2017}, year = {2017}, address = {Hawaii, U.S.A.} url = {http://sunw.csail.mit.edu/abstract/uberText.pdf} }