I am having issues with duplicate encoutners being created. For example I just created this encounter for a green sea turtle and I have two duplicate encounters. Usually the duplicate encounters only contain one image. See image of the screen shot of my encounters list. @Anastasia any reason this may be happening. It happens for almost every submission.
Moving this to its own topic while I investigate.
Here’s the original encounter the cloned image is from: Internet of Turtles
It looks like the two clones are a result of the image having multiple head annotations. Wildbooks assume that more than one of the same annotation type (like head or body) means that there are multiple animals in an image and automatically creates a new encounter for each additional annotation.
For some reason, IoT thinks it has detected 3 heads on this image. I’m going to flag this for the machine learning team to review since detections shouldn’t be misbehaving like this.
Here’s what I found out: Our algorithms weren’t trained on out of water, top-down shots. This is why the annotations on that photo appear misplaced.
We’ll need to get more of these and retrain the detector. In the meantime, you can delete the incorrect annotations and redraw them manually to see if you can get matches that way. Deleting the incorrect annotations will delete the cloned encounters they’re associated with.
Anastasia,
Thank you for the feedback, almost all of my submission have the top-down out of water shots provided. Can those help the detector and algorithm any? The link below goes to all my submissions through individual search.
1 Like
I’ve shared this with the machine learning team; thanks! I’ll let you know when I hear back with any updates.
There is currently no timeline on retraining the detector, but it looks like they’ll need have approximately 500 images using this viewpoint to get started on retraining.