Does adding noise or background training snips only affect our custom classifier classes predictions?

[Geematic]

We're training a custom classifier to find frogs - we're using append so we can find waterbirds in the same pass to save time.

We have recordings where running our classifier, we get a false positives for frog species - there is often a lot of background noise but unfortunately almost always with birds calling as well.

If we put these 3 sec samples with (unidentified) birds calling into our 'noise' training folder, will we affect the underlying bird predictions? or are the bird predictions independent because our classifier is appended.

(We run birdnet analyze with low min-confidence so we get everything included- we then filter birdnets CSVs by frog species classes and then by target bird species )

Also are 'background', 'silence' and 'noise' treated any differently to any other custom classes we train (other than not appearing in results)?

[Josef-Haupt]

The original "appended" classifier is not influenced by an of the training data for the custom classifier, only the appended classes might behave differently. We keep the original classes and confidences of birdnet as is, if you are using append mode. That means, if you are using a class that is already included in the original birdnet, with append mode, you will have two confidence scores for the same class, one from the original birdnet and one from your custom classifier. How you handle these scores is up to you, you can average them, use the max, really depends on your usecase.

We use "background", "noise" and "silence" all as the same non-event-class, they are treated the same.