Scene Text Detection in Natural Scene Images
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported License.
Task is available at: https://github.com/cs-chan/Total-Text-Dataset/tree/master/Dataset
Given a natural image, a scene text detection algorithm is expected to tell if there is any text instance in it, and return its precise location if it exists. Total-Text is different with all the existing scene text datasets like ICDAR2013, ICDAR2015, MSRA-TD500, etc., as it includes curved text instance in the dataset, which has close to zero existence in the mentioned datasets. The existence of curved text calls for tight polygon annotation format, which is different with all other existing annotation formats (i.e. axis-aligned bounding box, rotated bounding box, quadrilateral). Such a difference consequently requires new scene text detection solution.
Both DetEval and Pascal VOC detection evaluation protocols were implemented for this dataset. Modification was done to handle the interception calculation between polygons.
For the threshold values for DetEval (i.e. tp and tr), we recommend to set it as 0.6 and 0.7 respectively due to the use of polygon format ground-truth. The increase in precision threshold (tp) is to encourage a tighter binding result from detection algorithms, while the decrease in recall threshold (tr) is to relax the penalty towards the bias of a more flexible ground-truth format (higher vertex count in polygon).
All the images of Total-Text were seperated into a training set and a test set. Our implementations of the evaluation protocols are available on the GitHub page alongside the dataset.
The expected detection result should be arranged as (x0,y0,x1,y1, ...... xn,yn). There is no limitation on the length of your detection output.
No comments on this dataset yet.