How do I identify the postulated 600 object classes in OpenImages v. 6?


I have downloaded the OpenImages version 6 image sets (train, validation, test) from CVD Foundation as per instructions from the OpenImages homepage. The OpenImages homepage states that there are 600 boxable object classes in the dataset.
When I sift through the entire list of training set bounding box annotations from, it turns out 599 unique label names. So, that is one short of the supposed 600 classes.
If I download the boxable class names from (this is linked from the OpenImages version 6 page despite its apparent version 5 origin), this list contains 601 classes.

It seems neither of the above approaches identify the correct 600 classes, so how do I correctly identify the "official" 600 classes?

Thomas Arildsen

Posted 2021-02-24T13:31:43.393

Reputation: 131

No answers