In case a shop cannot be easily picked, it needs to be decided what image per category to choose. For example, the best image perspective for shoes is different than for blouses.

That is why we created a curated dataset for each category to rank an image according to its 'goodness'.

Each dataset looks like this:

 - folder:shoes
    - bad
    - good
    - hard_bad
    - hard_good
    - test

It is used for training and testing. The hard folders are those that are wrongly predicted *after* model training. This is an iterative process that continues as long as there are wrong predictions. In most cases, a 100% accuracy cannot be achieved.

The dataset is located on grapex: /home/picalike/projects/image_picker