Datasets

Download Datasets

An ImageTagger API was introduced in v0.3.5. See How to Download ImageTagger Datasets for more information.

Structure

Each top-level dataset is a named submodule, e.g. MLHelper.datasets.bitbots. These Submodules have datasets, in the case of bitbots they are split into TRAIN and TEST.

Bit-Bots

In order for Bit-Bots datasets to work you need to set the environment variable ROBO_AI_DATA:

export ROBO_AI_DATA=/path/to/root

All Bit-Bots datasets need to reside in that directory.

Available TRAIN datasets:

bitbots.TRAIN.LEIPZIG
bitbots.TRAIN.NAGOYA
bitbots.TRAIN.IRAN
bitbots.TRAIN.MONTREAL
bitbots.TRAIN.BITBOTSLAB
bitbots.TRAIN.CHALLENGE_2018
bitbots.TRAIN.ALL

Available TEST datasets:

bitbots.TEST.NAGOYA
bitbots.TEST.WOLVES
bitbots.TEST.IRAN
bitbots.TEST.BITBOTSLAB_CONCEALED
bitbots.TEST.CHALLENGE_2018
bitbots.TEST.ALL

Available TEST_NOISED datasets (i.e. noised versions of test images):

bitbots.TEST_NOISED.NAGOYA
bitbots.TEST_NOISED.WOLVES
bitbots.TEST_NOISED.REAL
bitbots.TEST_NOISED.CHALLENGE_2018
bitbots.TEST_NOISED.ALL

For an explanation of CHALLENGE_2018 see Towards Real-Time Ball Localization using CNNs by Speck et al.

Bit-Bots — Example

import MLHelper as H
from MLHelper.datasets.bitbots import BallDatasetHandler

# creates a dataset object with batch size 8 for alle datasets that were included in 'CHALLENGE 2018'
dat = H.ImgReader(BallDatasetHandler.TRAIN.CHALLENGE_2018, batch_size=8)