Does BirdNET perform detection before classification?

[shnchh]

Hi, just some quick clarifications:

1. During inference, does BirdNET perform detection to segment the audio clips into potential detected calls? Or does it split the whole audio into 3 sec segments and perform detection on all these segments.
2. When training a custom classifier, does BirdNET perform data augmentation on the new training set? If not, would it be recommended to do so?
3. I understand that BirdNET can append the new classifier's labels onto BirdNET's classifier. I am wondering how it works if there is a label/bird that already exists in BirdNET's classification layer labels, will the labels be deconflicted and merged before training?
4. What is the point of having a negative samples class in training since it won't be used for binary classification?

Thank you!