Is it normal that the matches take so long? Various members of my team have tried in on various machines and I from various different networks from both Indonesia and the US and it can take over 20 minutes or more for a match to process. I often quit the process losing patience. Are we doing something wrong?
Thank you for your help.
Ellie
If this is a bulk import report, send the spreadsheet to services@wildme.org with the email subject line matching your bug report
Yes, IOT uses Hotspotter and a suite of other computer vision tools used to isolate and orient the identifiable information.
Matches can take a while for several reasons. If you are starting multiple jobs at once, or there are many other users on the platform the queue for image analysis can delay results.
The other big factor is matching set size. If you are matching against many locations with many individuals and images, it can take longer. Right now looking at the queue for the last couple days I see blocks of ID jobs around 2-3 minutes, and blocks around 12-15 minutes. My guess would be that there is a big difference in the matching set size between those two user’s work.
If you send me an encounter you are working with I can take a peek at the candidate size and turn around time, but there likely won’t be any changes to make. I’d suggest starting several jobs at once and coming back to them later for results if the wait times are extended.
Hi Colin,
So I tried several as you suggested today by setting up 5 matches to work in the background. I set up them up this morning and over 12 hours later I am just getting an “attempting to fetch results” notification.
Here are the links the task numbers if that helps:
Hi, @E.Germanov !
When I click on all of those links, I see either “no matches found” or a list of match candidates. Maybe the 13th hour was the charm?
In regard to the processing time, it was likely taking a while and the page timed out, either on our server or your browser. The results page will indeed cease asking for results after 15 minutes so there is not a left open connection. My guess would be that if you closed the tab and came back to the results later through the menu in the corner of each image on the encounter page, it would shown as complete much earlier.
I looked again at the image analysis job history and see most jobs completing in around 2.5 minutes, with a handful of exceptions at >10 minutes. The time it takes for image analysis to transmit that result back to the browser is tiny.
If your jobs are taking a long time and the page is timing out, I would try being more specific with the location ID’s selected to match against. The results pages you linked show a matching set not limited by location- and this is a huge amount of data. The results pages quote around 26,000 possible candidates for a left side viewpoint world over.
I’ve found the problem here. The tasks with errors were due to manual annotation on the image
where turtle_green was used instead of turtle_green+head. Wildbook treated these as a body
annotations, which are not eligible for identification.
There was a config bug on the server side where we allowed jobs to be started for these, which subsequently failed. I’ve corrected the config, and the ‘start a new match’ for these annotations should not appear to be an option anymore.
To run ID on these images a new annotation will need to be added with the appropriate class, and the old one removed.
It appears this data was imported with an inactive location ID: Indonesia - Nusa Penida. The verbatim
location text can be anything, but the location ID field must be part of a list set by Wild Me in the database. When I ran a couple of these against just Indonesia data they completed successfully.
You can see the location ID’s we have active for Indonesia in the dropdown menu within the single encounter submission form. If there are a few additions to the Indonesia subcategory you would like to see, please advise us and we can add them. At that point they will be active for a photo ID matching set target.
The location Ids I see are Indian Ocean → Indonesia.
We also use Manta Matcher with the following locations based on our study locations: Indonesia - Komodo, Nusa Penida, Raja Ampat and Sangalaki which might be where the confusion comes from.
The two encounters still giving me trouble are:
92f0e179-b96e-4d05-b628-481a316e694a
d2285fcd-46bc-4088-b588-d4a91ec1d3cb
We generally just choose Indonesia and not any of the subsea categories (same as when running the matcher).
Hi, @E.Germanov !
For d2285fcd-46bc-4088-b588-d4a91ec1d3cb, it looks like the manual annotation is still for the body (and therefore not matchable). What happens if you delete that annotation and made a new one with class “…+head”?
As for 92f0e179-b96e-4d05-b628-481a316e694a, I’m not sure what’s going on with that one; the annotation’s class and the locationId both look good. We’ve recently deployed some bug fixes on IoT; could I ask you to delete that annotation, create a new one, and re-run matching, just to see what happens?
Hi Mark,
When I went to delete the annotation to d2285fcd-46bc-4088-b588-d4a91ec1d3cb, I also deleted the data image. Is this recoverable? I found the original so no drama, but it was not what I expected to happen.
For 92f0e179-b96e-4d05-b628-481a316e694a, I also deleted the annotation, but it did not delete the image.
