Dataset Card for Total-Text-Dataset

The Total-Text consists of 1555 images with more than 3 different text orientations: Horizontal, Multi-Oriented, and Curved

This is a FiftyOne dataset with 1555 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = fouh.load_from_hub("Voxel51/Total-Text-Dataset")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

• Curated by : Chee-Kheng Ch’ng, Chee Seng Chan, Cheng-Lin Liu
• Funded by : Fundamental Research Grant Scheme (FRGS) MoHE (Grant No. FP004-2016) and Postgraduate Research Grant (PPP) (Grant No. PG350-2016A).
• Language(s) (NLP): en
• License: bsd-3-clause

Dataset Sources

• Repository : https://github.com/cs-chan/Total-Text-Dataset
• Paper : https://arxiv.org/abs/1710.10400

Uses

• curved text detection problems

Dataset Structure

Name:        Total-Text-Dataset
Media type:  image
Num samples: 1555
Persistent:  False
Tags:        []
Sample fields:
    id:                     fiftyone.core.fields.ObjectIdField
    filepath:               fiftyone.core.fields.StringField
    tags:                   fiftyone.core.fields.ListField(fiftyone.core.fields.StringField)
    metadata:               fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.metadata.ImageMetadata)
    ground_truth_polylines: fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Polylines)
    ground_truth:           fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Detections)

The dataset has 2 splits: "Train" and "Test". Samples are tagged with their split.

Dataset Creation

Curation Rationale

At present, text orientation is not diverse enough in the existing scene text datasets. Specifically, curve-orientated text is largely out-numbered by horizontal and multi-oriented text, hence, it has received minimal attention from the community so far. Motivated by this phenomenon, the authors collected a new scene text dataset, Total-Text, which emphasized on text orientations diversity. It is the first relatively large scale scene text dataset that features three different text orientations: horizontal, multioriented, and curve-oriented.

Annotation process

Initial version of Total-Text’s polygon annotation was carried out with the mindset of covering text instances tightly with the least amount of vertices. As a result, the uncontrolled length of polygon vertices is not practical to train a regression network. The authors refined the Total-Text annotation using the following scheme. Apart from setting the number of polygon vertices to 10 (empirically, 10 vertices are found to be sufficient in covering all the word-level text instances tightly in our dataset), they used a guidance concept inspired by Curved scene text detection via transverse and longitudinal sequence connection paper by Liu, et al. which was introduced to remove human annotators’ bias and in turn producing a more consistent ground truth. The process for other annotations can be referred from paper.

The authors have mentioned in the paper that the human annotator was given the freedom to take a break whenever he feels like to, ensuring that he will not suffer from fatigue which in turn introduces bias to the experiment. Both time and annotation quality were measured internally (within the script) and individually to each image.

The authors have also proposed aided scene text detection annotation tool, T3, could help in providing a better scene text dataset in terms of quality and scale.

Who are the annotators?

Chee-Kheng Ch’ng, Chee Seng Chan, Cheng-Lin Liu and Chun Chet Ng

Citation

BibTeX:

@article{CK2019,
  author    = {Chee Kheng Ch’ng and
               Chee Seng Chan and
               Chenglin Liu},
  title     = {Total-Text: Towards Orientation Robustness in Scene Text Detection},
  journal   = {International Journal on Document Analysis and Recognition (IJDAR)},
  volume    = {23},
  pages     = {31-52},
  year      = {2020},
  doi       = {10.1007/s10032-019-00334-z},
}

Dataset Card Authors

Kishan Savant