Would this be able to detect true "unseen unknowns"?

[Zephyr69]

I've seen many open-set object detection projects that can detect "seen unknowns" (classes that are already present, although unlabeled, in the training pipeline).

But I'm looking to detect "unseen unknowns" (classes that do not exist in the image space of the dataset at all). Is this project capable of doing this, at least to certain capability?

[buxihuo]

I've seen many open-set object detection projects that can detect "seen unknowns" (classes that are already present, although unlabeled, in the training pipeline).

But I'm looking to detect "unseen unknowns" (classes that do not exist in the image space of the dataset at all). Is this project capable of doing this, at least to certain capability?

Hello, you can detect "unseen unknowns" objects. The ability to detect unknown objects mainly depends on the training set (the more categories of training set, the stronger the ability to predict unknown classes) and the parameter unknownconf (the greater the threshold, the less unknown objects will be predicted).
I have uploaded the code required for the complete detection process, but the training code will not be uploaded temporarily. If you need models trained under other public data sets, you can leave a message in the comments area.

[buxihuo]

The actual detection effect can be viewed by clicking the demo link on the home page