Tasks - ICDAR2017 Robust Reading Challenge on end-to-end recognition on the Google FSNS dataset
The single task associated with this challenge is a pure end-to-end text recognition task: users must "provide a full canonical transcription given the four views of the street sign".
A canonical transcription means that rather than containing the exact text observed in the image, it contains the text that would appear in a map. This normalization of the transcription mainly affects letter capitalization and discarding text that is not part of the street name as it can be observed in the figre. In this way, the task of transcribing the signs is not a basic OCR problem, but requires also some interpretation of what the sign means, not just its literal content.
During the training phase participants will have access to the public train, validation and test set of the FSNS datasets. Final evaluation will be done on a sequestered private test set that will be made avaliable at the beginning of the test phase.
Participants willing to participate in the competition with only a partial implementation of a complete end-to-end pipeline (for instance, with a method that performs only word localization or word recognition) will be allowed to plug their method into a modular end-to-end baseline method that will be made available by the beginning of the train phase.
FSNS: Link to the training set available
February 25, 2017 : Beginning of the train-phase
June 15, 2017: Beginning of the test-phase
June 30, 2017: End of the test-phase, final submissions.
November 10, 2017: Competition results.