Annotation issue

What Wildbook are you working in? Flukebook

What is the entire URL out of the browser, exactly where the error occurred? Flukebook

Can you describe what the issue is you’re experiencing?

The annotation for the whale is completely off - it is around water rather than the whale. Also, the annotations for all of the matching candidates are also very strange. They are mostly going the wrong way across the whales. Attached are some screen shots of a couple of the matching candidates (not sure if they are uploading properly).

Can you provide steps on how to reproduce what you’re experiencing? Just looking at the matching candidates for encounter 03ec272d.

If this is a bulk import report, send the spreadsheet to with the email subject line matching your bug report

Hi @jwaite, welcome!

The same issue was reported last week for belugas: Automated Annotation off

This is a result of training on “clean” data where pictures used are of ideal conditions and then presenting the model with more “messy” real-world images where conditions are less uniform. I recommend deleting the incorrect annotations and then manually adding new ones in the correct place.

Once you’re satisfied with the new annotation placement, you can choose “start a new match” from the image menu to re-run identification.

OK, thanks. I’ll talk to Christy about this.

1 Like

Hi there, Yes, we do understand that ‘off’ annotations on our bulk imports can occur, and that we should just delete, manually annotate and re-run match. However, the thing we thought was odd, was the annotations on the ‘historical’ or ‘training’ data were also off. So photos of whales that were already cropped and rotated and submitted for the training process are showing up in the matching process with odd annotations too. Can you help us understand that? Thanks!

Thanks for clarifying this. Let me see what I can find out for you.

Thanks for your patience! We’re in the process of recalibrating the annotations on these.

It looks like at some point, we went back to the historical data and re-ran the new detector on those images. The problem was that the new detector works best with the real-world examples, so when it saw the training set that was perfectly cropped and rotated, it couldn’t place annotations correctly because it was used to finding whales in the open sea, not already perfectly taking up the entire frame of the image. Basically it was the reverse problem of what we thought was occurring. :sweat_smile:

During the recalibration process, we’ll be cleaning up the bad annotations and re-run impacted beluga images through identification. This should hopefully resolve any similar match-related issues with belugas. I’ll let you know when that’s complete.

Great - glad to hear you found the problem. Thanks!

Glad to know we weren’t seeing things! Thanks for the explanation.

Just a quick update to let you know we’re still re-running IDs on belugas and the initial results are already looking much better than what we were seeing before. I hope to have another update by next week.

This is great to hear! Thanks.

1 Like

@Anastasia @jwaite , Just checking in on the progress/status of re-running annotations on the historical images. When that is complete we have new sightings we would like to run through Identification.

1 Like

Thanks for checking in! I just verified that they’re completed and you should be fully unblocked now. Feel free to upload your new sightings and review any previous encounters that you weren’t able to match.