labe issue

[DoublePan-Oh]

I am confused about the label problem. In your code, the label of the ground-truth is 0, which represent this is action, can I think that the lable output by the network is 0 for action, not 0 for background？ If the predicted label is 0, does it match the ground-truth?

[DoublePan-Oh]

Can the ground-truth label be modified, for example, to 1？1 means that the predicted label matches the ground-truth, otherwise there is no match. I now want to modify it into such a requirement, do you think it is feasible?

[sparkstj]

We take the last label entry as background. Since it's a binary classification problem, the first label entry is action. So yes, the label 0 represents action and the label 1 represents background. It's just a matter of notations, so it would be feasible to swap the labels for action and background.

[DoublePan-Oh]

This classification task should be a binary classification. According to the detr, it should be num_class=2 when setting num_class, but your code setting is num_class=1. This is where I don't quite understand.

[sparkstj]

In the COCO dataset and many other datasets, labels are expected to start at 1, so DETR leaves class 0 untouched and reserves max_id + 1 for no-object class. That's why they decide num_classes = max_id + 1.

In our case, we index classes from 0 to create more compact class embedding. We assign action to class 0, and the background/no-object class to class 1. That is to say, our max_id is 0 and therefore num_classes = 1.

You could also check this discussion for detailed explanations.