when I look at the classes it only shows the stuff classes in the annotations. about cocostuff10k HOT 7 CLOSED

Both stuff and thing labels are included in the JSON. Just open it in your editor and you will see that the category_id fields correspond to integers 1-182 from: https://raw.githubusercontent.com/nightrome/cocostuff/master/dataset/cocostuff-labels.txt

from cocostuff10k.

this is true for the json provided in this repo. However you refer to http://cocodataset.org/#stuff-challenge2017 for more images and in this json it starts at 1 with banner and ends with the category other representing everyting from the thing class.

from cocostuff10k.

Yes, but this repo does not have anything to do with the challenge, where all thing classes are merged into a single class. The code for that can be found here, where you can also find and modify the script that merged the stuff and things: https://github.com/nightrome/cocoapi/blob/master/PythonAPI/cocostuff/internalToCocoGTDemo.py

Note that banner has the index 92, so the thing indices (1-91) are still usable and you could just add the original thing annotations from COCO to the same file.

from cocostuff10k.

ok thank you. Do you know if the Img ID's in both sets match? Or even the AnnIds? Or do I have to match the filenames?

from cocostuff10k.

The image ids certainly match. The annotation ids are somewhat arbitrary. Just remove annotations with the new thing category id from the challenge annotations and add all annotations from COCO.

from cocostuff10k.

Ok that should be easy to do thanks! Do you know if the train val split is the same in the stuff and the coco set? Or should I combine them and split them after adding the annotations?

from cocostuff10k.

The split is the same, but since the current version does not include all 165k images from COCO yet, the training set of the challenge is a subset of all training images. Val is complete and hence identical.

from cocostuff10k.