We've spent quite some time getting CoralNet's vision backend ready for more advanced functionality. Slowly but surely, that functionality is finally taking shape!

Originally, a CoralNet source's images could only be automatically classified using that source's latest saved classifier. Then in 2020, we released the Deploy API, which allows any image with a public URL to be classified using any CoralNet classifier.

The API offers much more flexibility, but moderate programming knowledge and effort is required to actually use it. On top of that, if your images reside in a CoralNet source, you currently have to do webpage scraping to get those images' URLs so that you can pipe them into the API.

We wanted to make it much easier for classifiers to work across sources in CoralNet, since there are a broad range of use cases for this. For example:

  • You've created a well-trained classifier on a source, and you want to apply it to new data from a different dive/season/year or a different location, which you've chosen to upload to a different source.
  • You don't have enough data, time or resources to manually annotate enough images to create a well-trained classifier.
  • You want to use a publicly available classifier that was trained on carefully curated data by a trusted organization.
  • You want to standardize analysis across multiple sources.

So here's what's new: now, in the New Source or Edit Source form, you can choose the classifier that's used to classify the source's images. You can either have the source train and use its own classifiers, as before:

Or, you can have it use a specific classifier from any source, identified by its Global ID:

This can even be used to specify an older classifier from the same source.

You can find a classifier's Global ID by hovering over the corresponding plot point on the source's main page:

The source main page has also been rearranged somewhat, in an attempt to group classifier-related information more closely together:

Note that the source 'Affiliation' field is now displayed on this page, right under the 'Affiliation & Members' heading. Previously, affiliation was displayed on the source popups in the front page map, but not on the source main page due to an oversight.

Anyway, once you select a classifier, don't forget that you can check out your source's Jobs page to see how the vision backend processes are progressing:

There are two important restrictions on picking the classifier. First, if your source already has a labelset, it must exactly match the labelset of the classifier's source. Second, you can't choose classifiers from a private source which you don't have view access to.

You may choose to switch to a different classifier later. For example, you may want to switch to a newer classifier from the same source. The status text will say if there is a newer classifier available:

You can also always switch your source back to the 'Train' classifier mode, so that your source can train its own classifiers again, using its confirmed annotations as training data. As before, when a source is in Train mode, that source's latest saved classifier gets used for image classification.

Since this more advanced classifier functionality is still new, the details on how it works may still be subject to change, and there may be bugs to iron out. However, we're confident enough to announce it here and suggest trying it out!

(This post was updated on June 7, 2024 to provide some context/motivation for the new functionality, and to clarify that you can switch classifiers/modes again later.)