Large-scale Street View Text with Partial Labeling (ICDAR-2019 LSVT)

Research Tasks

End-to-end text spotting

2019-05-29 (v. 1)

Contact author

Yipeng Sun

Baidu Inc, Beijing, China


+86 10 56082834

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported License.


The main objective of this task is to detect and recognize every text instance in images in an end-to-end manner. Participants are asked to submit the locations of all the text lines in quadrangles or polygons along with the corresponding recognized results.


To compare the results of the end-to-end text spotting task more comprehensively, the submitted models will be evaluated in several aspects, they are : i) Normalized metric in terms of Normalized Edit Distance (N.E.D), and ii) Precision, Recall and F-score. Under the exactly matched criteria in F-score, a true positive text line means that the Levenshtein distance between the predicted result and the matched ground truth (IoU higher than 0.5) equals to 0.

The expected detection result is the locations of text lines in quadrangles or polygons and the corresponding recognized results for all the text instances in the image.


No comments on this dataset yet.

Add your comment

In order to comment on a dataset you need to be logged on
Register Now!


In order to rate this dataset you need to be logged on
Register Now!