Hi Matthias. Can I ask you why you put all those pip requirements in the requirements.txt file? I only needed some 3 requirements originally, I think.

Hi, sure! I basically kept installing dependencies until I was able to run the script in a new virtualenv. Then I just did pip freeze, so that's why it is very verbose. Probably tensorflow, future, scikit-image, scipy and matplotlib is enough. Would you like me to test?

Oh, and frankly I've just realized that my script does not download the results of the COCO dataset, so that evaluation fails. Let me fix that aswell.

Is matplotlib needed?

Also, you seem to have misunderstood what "mscoco_dir" is referring to. It's referring to the evaluation toolkit not the dataset. So this is wrong:
mscoco_dir = get_raw_input_data_dir("coco")

Can you please remove the backup text files (those that have ~ in their file name) as well?

In README you can say that the results folder has to be cleared before running the program.

Sorry I am taking so long to give feedback.