The idea behind ScrubberWatch is to detect the level of emissions of incoming ships in the Hamburg Harbour using Computer Vision. Essentially, we try to classify whether the Scrubber is turned on based on the smoke the ships are emitting. For detailed information on the use-case, please refer to the One-Pager and the small technical report.

ScrubberWatch is the winning project of the EIT Digital DeepHack Hamburg. We thank all the mentors, the organizers and especially Carsten Bullemer for the valuable discussions, ideas and feedback.

• Pitch slides
• One-Pager
• small technical report

The Dashboard is implemented in dash. To execute it, install the environment specified in the dashboard/requirements.txt via pip install -r requirements.txt. Then run python app.py within the dashboard directory.

Beware: the code for the dashboard is hackathon-quality!

• all relevant files are in subfolder smoke_detection

1. Download the dataset from kaggle and unpack it into smoke_detection. You should now have the following folder structure:

   ---smoke_detection/train/images/*.png

2. unzip the additional smoke data ship_smokes_fume_pollution.zip and link it to the train folder

   cd smoke_detection
   unzip ship_smokes_fume_pollution.zip
   mv ship_smokes_fume_pollution train/images

3. copy train_smoke.csv to train folder

   cp train_smoke.csv train

4. setup environment

   conda create --name scrubberwatch python=3.7
   pip install -r requirements.txt

5. train the recognition model

   python smoke_classification.py

6. run inference on input image

   python predict_smoke.py image.jpg

7. evaluation on validation set

   python smoke_validation.py

• please refer to the small technical report.

• The ship labelling tool is a simple jupyter notebook with ipywidgets. It can be launched with

cd smoke_detection
jupyter-lab
# open ship_labeling_tool.ipynb

• Maximillian Franz
• Sandro Braun
• Leander Kurscheidt

• Carsten Bullemer - Thank you for valuable advice towards this project idea.