For instance, the following image returns the JSON that follows.
It knows that the class of type “water” has a score of 0.83, which means that it is highly likely that that picture contains water. It also detected the activity of “swim” with a score of 0.83, which suggest that the picture contains the activity of swimming.
After parsing the JSON results, I created two files: one with a main classification such as ‘product’ or ‘activity’, and the other with subcategories such as ‘person’ or ‘swimming’.
The main categories are divided into fifty nine buckets, the two most common ones are ‘products’ (32% of total), and ‘people’ (24% of total).