Niklas Fiedler Hamburg Bit-Bots Department of Informatics University of Hamburg Export Format

Imagetagger:
An Open Source Online Platform for Collaborative Image Labeling

Niklas Fiedler
Hamburg Bit-Bots; Department of Informatics; University of Hamburg
Abstract To manage the large amount of images, image
The Imagetagger is a database with integrated tools sets are used. An image set consists of the im-
Export Format
to create and manage image data and related labels. It
was designed for the RoboCup to create training data ages it contains and additional information about
for neural networks and evaluation data for diverse ob- the whole set. This information includes the lo-
ject recognition methods. Therefore cooperative la- cation where the images got created. On the
beling of the same data set, flexible further use of the server side, the image sets are represented in di-
images and labels and the option to share the data had rectories (in the file system), allowing it to also
to be made possible.
work directly on the files. Image sets are owned
by teams, thus they are allowed to edit properties
and add or delete images.
Introduction
In 2016 Daniel Speck presented a neural net- Collaboration
work architecture to localize balls on images in
the RoboCup environment [1]. In the future
work section he proposed to increase the train-
ing data set for better results.
The lack of sufficient training data motivated
the development of the Imagetagger. As a re-
sult of providing the tool as online platform, the
exchange of image sets between teams is pro-
moted. The process of manual image labeling
has to be efficient enough to label large amounts
of images.
Image Labeling
Figure 3: In the team view, the image sets and members are listed
with the option to add team members and create new image sets.
The Imagetagger provides features to enable a

Figure 1: The labeling view consisting of the list of images, the image collaboration on two levels: The members of a
itself, a list of already existing labels and the annotation tool window. team and the teams among themselves. It of-
fers multiple options to export labels and up- or
The actual labeling procedure is designed to be download images which teams can use to easily
as intuitive and fast as possible. Therefore the share their data set.
UI is kept clean and simple with a flat menu Every team member is able to create and edit im-
structure. The most commonly used functions age sets owned by the team. Image sets owned
are mapped to keyboard shortcuts to optimize by the team can be set either private, to be only Figure 5: To create an export format, the user has to define the name
the labeling efficiency. A check box enables the accessible for team members, or public, which of the format, the types of labels and the base structure of the output.
option to keep the selection for the next image, allows all users to use the image set. This system
which is useful when labeling a series of similar offers a convenient way to share the training data To be useful for as many approaches as possible
images. The choice to label ”not in image” al- while providing the possibility to keep an image a flexible export format is needed. Every team
lows the user to explicitly label the absence of an set closed for public access when needed. has the possibility to export the labels in a format
object in an image. To support fast labeling, the compatible to their existing conventions. This is
next images get preloaded so even with a slow achieved by the export format creator, which of-
connection to the server the images are shown Automated Labeling fers a way to individually specify the format in
quickly. which annotations get exported into text files.
Already existing labels can be edited and deleted The automated labeling feature is currently work With the format barrier removed, the collabo-
directly in the labeling view. The bidirectional in progress. The Imagetagger offers a basis ration among teams is possible even with com-
continue buttons allow two users to traverse to enable automated labeling of images. As pletely different approaches and implementa-
through the image set towards each other. shown with a prototype, a large neural network tions of the use of the labels.
is able to detect objects, e.g. soccer balls, pre-
cise enough to create usable labels. To be able to Conclusion
Image Management distribute the labeling to multiple systems and to
separate the automated labeling from the website Currently the Imagetagger offers a way to man-
deployment, a REST-API could be used. age and create images data sets and labels. In
the near future an option to label automatically
Label Verification will be introduced to offer a way to produce a
lot of labels. Additionally it is possible to extend
the Imagetagger for labeling other types of data,
because it’s easily extendable and open source.
The code is available on GitHub:
https://github.com/bit-bots/imagetagger
References
[1] Daniel Speck. Ball localization for robocup soccer using convolu-
Figure 4: The image set management view showing the content of the tional neural networks, 2016.
image set, status information, a download script, an editor for meta
data and an upload area. Acknowledgments.
Labels can be verified by users to increase the Thanks to the RoboCup team Hamburg Bit-Bots. Thanks to the Server
AG and thanks to Marc Bestmann, Rebecca Glaser, Jonas Hagge,
confidence in them. While manually created la- Jennifer Meyer, Daniel Speck and Pascal Wichmann for their
bels get verified by their creator or last editor cooperation and help.
by default, the automatically created labels are
unverified on creation. These labels need to get
Figure 2: The image set management view showing the list of pictures verified by a user. The amount of verifications
included in the image set, status information, a download script, an can be used to filter the labels.
editor for meta data and an upload area.

Niklas Fiedler Hamburg Bit-Bots Department of Informatics University of Hamburg Export Format

Uploaded by

Copyright:

Available Formats

Niklas Fiedler Hamburg Bit-Bots Department of Informatics University of Hamburg Export Format

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Niklas Fiedler Hamburg Bit-Bots Department of Informatics University of Hamburg Export Format

Uploaded by

Copyright:

Available Formats

Imagetagger:

An Open Source Online Platform for Collaborative Image Labeling

The Imagetagger provides features to enable a

You might also like