Hamburg Bit-Bots Ball Dataset 2018

The Hamburg Bit-Bots Ball Dataset 2018 offers a way to train, benchmark and compare ball localization methods out of the RoboCup Soccer environment. Please see the corresponding (acceptance pending) paper Towards Real-Time Ball Localization using CNNs, 22nd RoboCup International Symposium, 2018.

Information and code about the models is shared on the GitHub repository: Towards Real-Time Ball Localization using CNNs.

The test and training data is provided via the Bit-Bots Imagetagger. It allows the creation of customized label export formats. For more information on the tool, see this poster or its GitHub page.


You can download everything at once: we supply a zip file with all datasets (including negative dataset) on our cloud page: Hamburg Bit-Bots Ball Dataset 2018. Labels are included in txt-files (our format). If you prefer to get the labels in your own custom format you can do so by clicking on the imagesets below, specifying your export format and exporting the labels yourself with our Imagetagger.

Training datasets (22927 images and labels):

Test datasets (2177 images and labels):

As error measurement, we propose using the accuracy rates within an error radius of 3, 5 and 10 pixels from the center point.