Hi team- The following encounters are showing a “no matchable detection” error in the hamburger menu:
They don’t appear to have bounding boxes drawn, and I feel like maybe we talked about this in a meeting a while back, and about the possibility of eventually being able to draw the bounding boxes ourselves when this happens, but I don’t remember if we talked about how to troubleshoot in the meantime. Thanks for your advice!
-Chrissy
Hi @ChrissyTustison!
Let me work through everything you presented:
First, we checked and these encounters did fail on matching. We are kicking those off again and will let you know when it’s resolved.
Second, we are actively working to address self-service on this. There is work is currently being developed for the MantaMatcher platform to allow for an update to their matching algorithm that allows users to draw their own bounding boxes, and the changes will cascade to other Wildbooks once completed. For your reference, this work is being tracked under ticket WB-285 and is in active development, so should be available in the next month or so.
Thanks so much!
Tanya
Thanks, Tanya! That sounds great. I’ll stay tuned for the new matching results for those encounters, and good to know about the features under development for MantaMatcher. Looking forward to hearing more about that!
-Chrissy
Hi Chrissy,
I’ve run those annotations through the detection platform again and retrieved annotations. This does not automatically kick off identification, but I’ve sent off a couple of the images and recieved results. You can try the others with the usual ‘start another match’ button. I wasn’t able to see why detection initially failed, but there are no errors in the log which makes me suspect a case of absolutely perfect timing when the image analysis software was going through an update. I’ll keep checking the logs, and do let me know of you see this again.
Thanks!
Thanks Colin! I was able to get good matching results for both of those encounters. Now I’m seeing separate encounters for the photos that have head bounding boxes (the ones I matched on) and photos that have body bounding boxes- should I add the ones that have body bounding boxes to the encounters where I matched on the heads, or delete them? Not sure what the system will like me to do!
This is one group of encounters (NSW-0071):
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=c43dd12c-55fd-4203-9732-ac2e74c10fc6
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=5f605336-ae14-42a1-a1b5-829a1512719a
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=2493267b-6b19-40ef-964a-550fd150873a
And this is the other (NSW-0060):
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=d6642640-e3f7-46b0-ad06-f9957ffae7f9
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=cc86e5c5-1203-4928-ab27-d2016305e4be
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=5c0fba9e-a663-4693-b7ea-bcc5ec6a785d
I also noticed for NSW-0060 that there was an IA error message for the second photo:
https://wildbook.dragonsearch.org/iaResults.jsp?taskId=8e3659b8-da5d-4e81-8e40-6f722ef07288
Thanks again for your help!
-Chrissy
Hi again- ran into some other encounters with no matchable detection!
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=129a08ec-b274-4be9-b31d-f403b9ec340f
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=55b9ea55-6f10-48e7-a4d8-811e6326d639
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=9d092a77-db34-4c94-a702-47846dc8fd97
This one has no matchable detections and has also been split into 2 encounters with one for the head detection and one for the body detection:
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=3a3e7a99-915d-456c-8765-4d6b2152acb5
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=4b0be69b-0093-4699-a50e-07995c76d76f
Thanks for your help!
-Chrissy
Hi Chrissy,
I’ll open this up again with the issue tracking ticket WB-402 and be in touch with more information.
Thanks!
Hello,
I’ve gone through all the encounters listed. There were a couple annotations that needed to be reset to matchable classes, and I this should now be done for all of the head bounding boxes.
I’ve also found a bug that may have led to this occurring and it should be solved for future submissions.
The image with an IA error message looks to be processing normally now. Please visit the new results from the encounter page, as the above link will still bring you to the old results.
For the split encounters, unfortunately there is not a great automated way to condense them back into one. The two solutions are to delete the encounter with only the body annotation and continue to use the one with the head annotation, or to delete both and resubmit. Since you only use the head bounding box and the whole image is still present in both I recommend the first option.
I’ll wait to mark this as resolved until you can review and comment. Thanks!
Thanks, Colin!
I can see the matches now, but some of the lefts and rights of the same dragon appear to have been split, and the info entered under “additional comments” has disappeared… So I’m going to need to go back through the original files and re-work all of that, then I can assign the dragons and let you know if it all works!
In the meantime, I got a new error today when assigning an ID:
Failure: Encounter 3a3e7a99-915d-456c-8765-4d6b2152acb5 was NOT added to NSW-0073. Error unknown.
It was for this encounter:
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=3a3e7a99-915d-456c-8765-4d6b2152acb5
I tried refreshing, logging in and out, searching for that ID to see if it was already assigned somewhere else, but no luck. Any advice is much appreciated!
Thanks,
Chrissy
Hello,
It looks like this image has been submitted in another encounter with another individual name:
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=3c435dfa-8dff-496f-81db-282d566f2e17
This message you saw was a stop to prevent those naming conflict IA errors we have run into a couple times.
You can see when an image has a duplicate entry on the match results page:
https://wildbook.dragonsearch.org/iaResults.jsp?taskId=658db66d-7f96-4fa8-ab98-05537378ea0d
On the left with the target image you can see “alternate references” and a small link to the other encounter. (Remember the above link will break if a new matching job is run for the image)
Hope that helps!
Thanks for that last message of advice, Colin! Good to be reminded how to see the duplicate entry issue.
I tried deleting and resubmitting one of the encounters that was previously split up, and again submitted as one encounter (left and right of the same dragon, same dive/date/location), but it processed as 3 separate encounters:
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=54a7d439-d329-4137-8aa6-2b797fef266c
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=bb710caf-ea22-4826-b3b5-c690738f909d
https://wildbook.dragonsearch.org/encounters/encounter.jsp?number=791ee64a-dd91-4081-96dc-cce6dd3b11b2
In addition to splitting the photos up, there are bounding boxes drawn, but again the hamburger menu says “no matchable detection”. Thanks for your advice on this!
-Chrissy
Hi Chrissy,
Sorry for the trouble. I’ll look into the issue again. It looks like the error has been fixed for new images going forward, but some images that encountered the error previously are being reset to the incorrect class. I’ll keep this issue under review for these annotations and keep you updated.
Thanks.
Colin, I figured something out! I’ve been working from a stash of photos shared with us by a dive organization, and I’ve now realized there are duplicates in the photo set. For example, in the encounters I linked in my last comment, that is the same dragon as NSW-0059, which I’m guessing has something to do with why the system is reacting strangely to the submission?
Those photos were repeated in two different folders of the shared photo set, with different names, dates, etc, and I didn’t pick up on that at first. Just wanted to let you know that I finally figured that out, and I’ll delete those unapproved encounters unless you want to use them to study anything about the system’s response on your end? Thanks again for your consistent help processing photos!
Best,
Chrissy
Hi Chrissy,
Yes, that could definitely be a source of error! Any time a matching set has two identical images with different names, the job will stop in it’s tracks.
Go ahead and delete those encounters. If you have any others that did not process send them my way, I want to make sure that the previous condition we addressed isn’t still occurring.
Thanks!
Okay thanks! Deleted those and will keep you updated on any other issues.
Hi Colin- a few new encounters with “no matchable detection error” (for each of these, most photos processed correctly but some did not)~
SeadragonSearch.org (left-side head shot missing annotation)
SeadragonSearch.org (left-side head shot missing annotation)
SeadragonSearch.org (left-side head shot missing annotation)
SeadragonSearch.org (right-side head shot missing annotation)
Thank you!
Hi Chrissy,
I’ve gone through a couple steps Re: the missed detections at the beginning of the thread, and these more recent ones.
Jason and I worked to correct any incorrectly assigned classes, species or missing parts in the Image analysis server and just refreshed many using the most current model.
I’ve also queried through the Wildbook database and looked for things that might have been amiss, like missing species name or an asset without a head detection. All of these I gathered up and sent back to detection manually to make sure they had the proper class and labels set.
I found 4 annotations that shifted slightly, but none were the ones in your examples. The images without matchable detections in the above encounters I believe are cases where the model ran correctly but could not isolate a head annotation from the body. Three of the images are of a mostly obscured animal with only the head exposed, and they were classified as whole animal annotations. The fourth is in murky water, a small portion of the image and with significant rotation- no detection was made on that image. These images just seem to be outside of what can be recognized by the current model.
Sorry the result is not more satisfying regarding those 4 images, but the process has functioned as a great integrity check on the other data and classifications, and we can be absolutely sure the issue that caused the missed detections at the top of this thread is no longer active.
Talk to you soon.