Can you describe what the issue is you’re experiencing?
For several encounters we’ve noticed that the photos are rotated 90 degrees when viewed using the “inspect” tool. Not sure if this is supposed to happen, as it makes them difficult to inspect!
Can you provide steps on how to reproduce what you’re experiencing?
Possibly by submitting a new encounter and then checking to see what happens when you use the inspect tool to examine match results.
If this is a bulk import report, send the spreadsheet to services@wildme.org with the email subject line matching your bug report
This is HotSpotter rendering potential matches (via the “inspect” button) according to a standard orientation predicted by the orientation ML network. The ML orientation network rotates them according to where the dotted line is placed during the annotation process. For seadragons, it looks like the dotted line was placed at the front of the snout, and this is being displayed as the “top” of the image.
So, it looks like the system is working as designed. I recommend putting in a feature request to re-train the orientation network after fixing the annotations (i.e. Wild Me scripting a fix to rotate -90 or 90 degrees depending on the viewpoint).
Until then, there are browser plugins that can be used to rotate the HotSpotter image for review if it is confusing.
This is an example for Chrome (I have not tested it though):
The orientation network is new, so it may cause us to review how we annotated data for other species as well.
Are both of these annotations new annotations since the orientation network was turned on? One thing to keep in mind is that we did not re-run old annotations through orientation since you had made decisions based on those, and we don’t like to touch data on which decisions have already been made.
The orientation network will only be properly aligning pairs of new annotations in the HotSpotter inspect feature. New-old matches may appear misaligned since older annotations don’t have a predicted rotation.
Hi @jason- thanks for those replies! I think we have another issue as well though- we’ve noticed that some obvious matches are no longer being made, for example this one:
This weedy is NSW-0179 and has lots of matches in the database, but none of those were suggested in this round of match results, which makes us hesitant to assign any matches if the system is missing the correct ones. Maybe we should turn the orientation network off and see if things go back to normal?
I turned it off. The orientation network will improve matching performance overall…but because of standardization of annotations.
We’re discussing internally. It may be that use of an orientation network should be an “all or nothing” affair, with all orientations standardized by it…rather than a mixed use of the network like we’re doing here (old annotations unoriented while new annotations are oriented).
Please keep in mind that annotations created since we added the network will still have their rotations. Old annotations, and annotations added as of today will not.
Thanks, @jason ! I think we will delete and re-submit encounters with the rotated annotations, to keep everything consistent going forward. It seems like a great tool, I think we just didn’t realize how much it would impact matching against all the data we already have.