What is the class id for background class? about detr HOT 9 CLOSED

@raviv DETR uses the last value for the class.
So if you have 10 classes, you should set num_classes to 10 and class index 10 will be the background,

EDIT: corrected the value for num_classes, the no-object class is added in the model definition

from detr.

Hi,

No-object label is num_classes + 1 in DETR, so the last element in the prediction.

There is no need for the dataset to return no-object classes (as it return only existing objects). The only place where the notion of the no-object position is taken into account in the loss and postprocessor.

Let us know if you have further questions

from detr.

@fmassa usually when training 10 classes, 0 is for bacgkround and 1-10 are actual classes.
Do you mean that here 0-9 should be the classes and 10 is for background?
Also, when constructing the DETR class should num_classes param be 10 or 11?
Thanks.

from detr.

@fmassa
I am bit confused with your answer after the EDIT above. Does the first statement hold?
"So if you have 10 classes, you should set num_classes to 10 and class index 10 will be the background,"

from detr.

@hariSky Yes, the mode build with real num_class N that no need for background class. But the model itself set with N+1. The N+1 is no object.

from detr.

@fmassa fmassa Hi , can DETR train images that without any object on it ? I mean the negtive data .

from detr.

@fmassa same question as @zhishao . For an image without an object(negative example), how do we label it in COCO format? Do we label the annotations bbox as all zeros? Or we do not even label it, since there's an extra class built in to handle no-object? Thanks!

from detr.

I want to ask the same question, I want to compare YOLO and DETR. I found that in YOLO, an image without an object(background images) just leave it alone. YOLO will take care of it. Does DETR do the same way, for the background images?

from detr.

I wanted to know if there's a clarity on this. I've total of 3 object classes, Since DETR outputs num_classes + 1. In coco format how should I even label the negative samples, should I keep the category_id=4 and for bbox=[0,0,0,0]. Any help would be greatly appreciated. Thanks in advance !

from detr.