When I refreshed the browser page, the annotation disappeared.
What did you expect to happen? I would expect that an annotated image would allow us to run matching. And I wouldn’t expect an annotation to disappear from a media asset.
I’ve taken a look at this image and found the issue. While detection completed successfully, it was still marked as processing internally and therefore not allowed to go to identification.
There is a lot of system logging water under the bridge since this import on 2020-06-22, but I did uncover something- this image has been submitted three different times in import. Once it was unassigned and twice it was submitted with different names.
The two identical images with different assigned marked individuals were part of the same bulk import, and the images are simply named differently, and part of different encounter rows.
It’s very important to make sure identical images (therefore the resulting annotation of the animal) don’t have conflicting individual names. This can cause errors in the image analysis pipeline.
I’ve manually set the affected annotation’s detection status, but this naming conflict must be addressed before identification takes place.
We are aware of the multiple submission of images - a result of the manual matching and curation process in the past on legacy data sets.
Since we are also now uploading raw data from citizen scientist sightings and where some of these images were used by the researcher to create ID kits for Individuals - we are expecting to see duplicates at that level also - albeit with a filename difference.
Other than doing what we are doing currently, that is - to rely on the system to find these for us and allow us to “correct” through the matching process and/or ID merges, is there any other way of determining which images fall into these categories ?
Short answer is no, we do not have any automated tools capable of flagging duplicate images with conflicting names for the user to reconcile prior to identification.
The system itself is unaware of duplication until the image is sent to detection and an identical
bounding box is returned. Even then it does not cause any problems unless conflicting names are
assigned.
We have had some discussion on ways of providing lists of conflicting ID’s to users, but it’s not a small amount of work unfortunately and would not prevent the halt to identification prior to resolution.
The best course right now is to try and prevent this happening- if you suspect that a duplicate image may have been used in encounters assigned to different individuals try to crop appropriately it or hold it back from the encounter with more images.
