We’ve retrained detection (bounding boxes and their rotation) and viewpoint labeling and need your help with the next steps.
There’s two successes/failures we need to test:
Are bounding boxes generally better, including new and old perspective
If yes to the above, are viewpoint labels as-good-as or better than before?
There are some trade-offs between the top three candidates for our labeler (such as certain viewpoints getting labeled correctly more often than others) and we’d like to hear from our turtle researchers if you want to participate in these experiments to better evaluate its performance, especially if you’re working with less common viewpoints.
Let us know if you’re interested in helping with the the detection experiments by posting here!
Happy to help if further tests are needed for individual uploads - ORP team has also noticed several issues with IA Class and orientation assignments for automatic annotations.
@ORP-StephanieK asked us to hold off on starting the detection experiments until the bug that prevented users from editing annotations was fixed. We rolled that out yesterday evening, so we’ll pick up that discussion again this week. Thanks for your patience.
I hoped that the detector would have gotten set up by this weekend, but I’ll need to follow up on it next week. Here’s what I know so far:
The test model will use the old viewpoint labeler with the new detector. In order to test it, you’ll use the taxonomy Chelonia fictus. This allows you to compare the old detector performance (using the correct taxonomy for your turtle) alongside the new one (using the fictional taxonomy).
We’re looking for feedback on annotation accuracy, if the correct viewpoint labels are applied (left, right, up, down, etc.), and matchability.
Tagging @jason in case you happen to set this up before I’m back online.
