num_classes about up-detr

[dddzg]

I think you are right.
But, if you set num_classes to 1 you should change

up-detr/datasets/selfdet.py

Line 73 in be066df

labels.append(1)

this line to 0.
Because, the last class is designed for no-object category.

[rgbd-zml]

I think you are right. But, if you set num_classes to 1 you should change

up-detr/datasets/selfdet.py

Line 73 in be066df

labels.append(1)

this line to 0.
Because, the last class is designed for no-object category.

Thanks for your reply!
If the num_classes is set to 1, this means that the label-0 represents the object class and the label-1 represents the non-object category in the classification cost?

[dddzg]

Yes. So, in our code, label-0 is never used.

[rgbd-zml]

Yes. So, in our code, label-0 is never used.

Hi, I have also a confusion about the paper. There is a explanation in up-detr paper:

In this paper , label-1 represents the object category(matching). There are some differences between the code and the paper. I'm not sure when num_classes is set to 1, which one of label-0 and label-1 represents object category(matching)?

[dddzg]

I am sorry for the inconsistency.
In our paper, 0 represents no-obj, 1 represents obj.
In our code, 1 represents obj, 2 represents no-obj. (0 is nerver used, and there is 3 class num in total)

up-detr/models/detr.py

Line 40 in be066df

self.class_embed = nn.Linear(hidden_dim, num_classes + 1)

Check here for the loss calculation.

up-detr/models/detr.py

Lines 124 to 126 in be066df

	target_classes = torch.full(src_logits.shape[:2], self.num_classes,
	dtype=torch.int64, device=src_logits.device)
	target_classes[idx] = target_classes_o

If you set num classes to 1,
0 represents obj, 1 represents no-obj.

The last class always represents the no-obj in our code.

[rgbd-zml]

I get it! Thank you very much!