Hi, I know that the MiewID detection will create clones, but I wondered if it can be disabled on this species Carcharias taurus since it just creates more noise than help. It clones faces of other sharks, fins, tails and its really unhelpful for this species. Can you look into whether it can be disabled?
When I have a photo of multiple sharks I already try my best to upload multiple times so I really dont want or need clones.
MiewID is just the matching AI model. It doesn’t clone encounters or create the bounding boxes. That is handled by a custom AI detector model for the species.
We can definitely retune it or turn the detector off if it is causing too many false positives and clones. Can you send some examples so I can look at whether retuning, retraining, or disabling it would be better?
Since Im the main one using this species in the database, I can say its almost always unhelpful clones, but happy for you to have a look at it. The clones are nearly always of Fins/tails/secondary sharks in the background. But for this species, unless we can see the spots on the shark, we will not be able to identify the shark and so I am concerned that clones are just flooding the database with more bad images.
If you check these out they are all examples of unnecessary/unhelpful clones:
I found three things in these examples:
-a lot of secondary and tertiary shark annotations
-a few fins being annotated inside the body
-at least one false positive background annotation
The secondary and tertiary sharks are annotated by design. The model is trained to draw bounding boxes around every sand tiger in the image and clone the encounter for each additional shark it finds. This is by design. Feel free to delete the additional encounters if you do not want them.
The fins being annotated inside the body annotation is undesirable. I have increased the non-maximum suppression (NMS) for the detector, which influences how overlapping bounding boxes occur. We should see fewer of these moving forward, and if it keeps happening, there is more room to increase the value. Too high, and we begin missing sharks. Too low and we get overlapping bounding boxes. We’ll iterate until we find the happy medium.
The false positive bounding box is simple model error. Feel free to delete those encounters when they happen. I’ll ask the ML team for some tuning adjustment if we see more of those moving forward. Feel free to keep sending them so we get a signal of how to tweak them.