A image-labelling platform written in JQuery/Bootstrap and python 3/Flask which persist annotations in a MongoDB database.
OpenMTurk helps you put together a supervised-learning image data-set. OpenMTurk provides you with your own in-house image-labelling platform.
- Install OpenMTurk
$ git clone https://github.com/HichameMoriceau/OpenMTurk.git && cd OpenMTurk/$ make install
- Customize the JSON config file to suit your classes/needs
$ gedit config.json
- Build
$ make
- Annotate
- By default the tool runs here:http://127.0.0.1:5000
Out of the box, OpenTurk comes with the following labelling features:
- Image class
- Image orientation
- Bounding-boxes, textboxes and lines
By default, the server expects images in the static/notes_photos/ directory. However, you may change this by overriding the IMG_DIRECTORY variable in the Makefile. You will be interested in doing so if you have lots of images.
Update the project/config.json to add, remove or rename any button. This file defines the buttons for categories, orientations and bbs (bounding-boxes).
For bounding-boxes there are 3 types of drawing buttons:
- box: a simple bounding box,
- textbox: bounding box which then ask you for the text inside the box (the textarea will become writeable),
- line: a simple line
{
"categories": [
"Form",
"Receipt",
"Letter",
"CV"
],
"bbs": [
[
"Text",
"box"
],
[
"Word",
"textbox"
],
[
"Image",
"box"
],
[
"Separation",
"line"
]
],
"orientations": [
"up",
"left",
"down",
"right"
]
}OpenMTurk persists all annotations to MongoDB. You can retrieve all of them by clicking the "Download all labels" button at the bottom of the main (index.html) page.
Note that OpenMTurk dumps a snapshot the database every 10 annotations into a file named labels_backup.json.
We didn't want to close the potential for having your images annotated by a team of data entry specialists. With this tool, you can build, deploy this site and hire annotators to label your images.
In case of an unexpected server or database crash, OpenTurk dumps all persisted annotations into a labels_backup.json file every 10 annotations by default.
Available projects/to-do list:
- add support for segmentation labels (masks or polygons)
- add a "Getting your images labelled by external annotators" section to the
README.md - deploy an example website
- add an Active Learning strategy to improve labelling efficiency
- add a random sampling strategy (instead of sorted return)
Other ideas? Let me know.
- You found a bug? You need an important feature? Please raise an issue!
- You need help getting started? Something isn't clear? Let me know.
OpenMTurk is not the only image-labelling tool out there:
Copyright (c) 2017 Hichame Moriceau & Larissa Tijms, the content of this repository is bound by an MIT License. Therefore, you may use the content of this repository for commercial and non-commercial purposes.