where there is a big foreground blocking element (like a truck in these examples) in an image, detection is significantly degraded for any target species in the background.
This is something that would likely involve an update to our ML practices, so I’m gonna break this one out into a feature request and have JP look at it.
Yes, as these types of issues are marked, the system has the potential to improve over time. We will need to mark these with corrected boxes and then retrain the detection model.
I ran a few different configurations against the image to see if the detector would place a box on that background animal and, once it finally did, the box scored very low (around 15% confidence when we normally require 48% confidence). The system is optimized to find Annotations of Interest (AoIs) and that background animal is very challenging from an ID perspective, so it was not prioritized during the initial training process.
Hi @parham, hope you’re keeping well. Thanks for the info, it’s interesting and makes perfect sense. The truck example will be less common for us, I believe, than scenarios where a tree or termite mound partially occludes an animal. We will definitely box all animals that get missed in detection so we’ll have lots of material to re-train with.
thanks
M & @PaulK