Should `MODEL.ROI_HEADS.NUM_CLASSES` include background?

[FraPochetti]

Hello everybody,
first of all, thanks a lot for building this awesome library!

I have checked the docs and could not find a clear answer to my question, so asking here.
I am currently using a mask_rcnn model and I am not sure whether MODEL.ROI_HEADS.NUM_CLASSES should include background or not (e.g. if my dataset includes only dogs, NUM_CLASSES should be 2).

This is not specified in the relevant class docstring and, given torchvision makes it explicit, I am a bit confused.

Can you please provide some guidance?
Thanks

[huangfeng95]

I have the same doubt

[ppwwyyxx]

The config reference https://detectron2.readthedocs.io/modules/config.html#config-references has answered this: "Number of foreground classes"

We will add it to the class docstring as well.

[FraPochetti]

I am more than happy to open a PR for this!

@ppwwyyxx thanks a lot for the prompt reply.
I just want to make sure I get this right.

"Number of foreground classes" means that it should include background, so in my example above "if my dataset includes only dogs, NUM_CLASSES should be 2".
Is my understanding correct?

[ppwwyyxx]

"Number of foreground classes" means that it should include background

I'm curious why this is understood this way? foreground and background are two opposite word, so I would interpret it that "Foreground classes" do not include "background"

[FraPochetti]

You are right.
The reason is that I misinterpreted foreground and background as synonyms.
Just bad English on my side I guess. My apologies.

I think it would probably be an idea to be more specific in the docs, explicitly mentioning number of foreground classes, e.g. not including background.
Some sort of "explain it to me like I am 5" to avoid any other bad English speakers to get confused in the future.

Thanks a lot for the reply!
Do you want me to open a PR about this?

[binhere]

The config reference https://detectron2.readthedocs.io/modules/config.html#config-references has answered this: "Number of foreground classes"

We will add it to the class docstring as well.

Hi @ppwwyyxx

Thank you for your instruction, It is very helpful; however, this document about dataset https://detectron2.readthedocs.io/en/latest/tutorials/datasets.html#standard-dataset-dicts also says:

"category_id (int, required): an integer in the range [0, num_categories-1] representing the category label. The value num_categories is reserved to represent the “background” category, if applicable."

Actually, I am confused about that statement "The value num_categories is reserved to represent the “background” category, if applicable."
How do you understand this ?

In my case, when I create a dataset on Roboflow and export it as COCO segmentation format, my dataset (has only 1 class "shelf" ) always includes class 0 for background. Here is my json file about categories field:

categories": [
{
"id": 0,
"name": "0",
"supercategory": "none"
},
{
"id": 1,
"name": "shelf",
"supercategory": "0"
}
]

I am now struggling with these things:

MetadataCatalog.get("train").thing_classes should include background? for example, ["0", "shelf"] ?
or
how can I adjust my json format to be compatible with detectron2 expects ?

However, follow the explicit document that you refered, I think my following code is correct:
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 1 # for foreground class "shelf" only