What Wildbook are you working in? Mantamatcher - Bull rays
What is the entire URL out of the browser, exactly where the error occurred? n.a
Can you describe what the issue is you’re experiencing?
Hello,
I’m a bit concerned at the lack of matching with the bull rays. I’ve tried matching repeated Individuals I know have been submitted but they did not come up as a match. For example ones named Jupiter, Ghary or Grumps.
I’ve tried making a project so I could start building a catalogue and tried matching from there instead than from the encounter page. But this was also unsuccessful.
I am unsure If I’ve done something wrong? Maybe skipped a step? I thought maybe it was beacuse the system had to be trained but when I went into the maunal it said that the hotspotter algorithm did not have to be trained.
The gap of time between when we were included in mantamatcher and when we started using it was quite large and I’m worried we’ve just gone off track.
I have investigated some other similar problems in the forums but they don’t seem applicable. I’ve tried the wiki as well, but I have not been able to sort it out.
Do you have any advice on how to proceed?
Thank you and sorry for any confusion.
Can you provide steps on how to reproduce what you’re experiencing?
This happens when I try to match from the encounter page, through the box “start matching” or from the manage projects tab.
Thanks for providing these links. It looks like you’re doing everything correctly. MantaMatcher and the HotSpotter algorithm also seem to be working correctly.
After reviewing the links, I think a lot more data needs to be processed before we understand how well or poorly HotSpotter is going to work for this species. Currently, there are only two individuals marked in the database across 100+ encounters, and for the example of of Ghary_001, some Encounters match:
When you have more individuals curated, we can create a graph of the top-1, top-5, and top-12 match results and get a comprehensive understanding of how well we are matching. Currently, there is very little data.
Two things come to mind:
HotSpotter aggregates scores across Encounters and images of the same individual, boosting its match results. Ghary_001 (and Grumps and other individuals you add) should become more matchable over time because the scores of multiple, matching photos will be summed and boost its rank in the results. The scores are liked by the animal’s ID.
is likely due to the lack of visual contrast (HotSpotter is a texture matcher) compared to the other example photos of Ghary_001. But now that both sets of photos are under the ID Ghary_001, we’re less likely to miss the match on future photos as we now have diverse representations of the same pattern.
Make sure you’re aware what you’re matching against regarding location. MantaMatcher provides the ability to filter based on locationID:
If an Encounter has locationID “Malta”, it will automatically select Malta and the subsets in the “Choose criteria to match against” dialog box. But if it is set to location ID “Golden Bay”, like this:
then it is only going to try by default to match against Golden Bay, missing other potential matches from other study sites. To make sure you are matching the subsites, select them all or click “select none” at the top to match against all Encounters for the species.
This is a very cool species, and we look forward to seeing how HotSpotter performs as more data is curated.
Since this last post more individuals have been added and some matches made - although primarily visual. For example this individual was sighted 3 times: MantaMatcher
He was matched visually, all pictures are good.
When I try to “start a match”, it does not find the other sightings. Is this because they are already matched before?
There are only 8 individuals in MantaMatcher with >=2 encounters, and only three with 3 encounters. It’s really too early to understand if HotSpotter is or is not performing well for the species. We can also find positive examples of matching, like this one:
Since your individuals are moving between these study sites, I recommend either unchecking the box, which will match against all Encounters of the species, or at least checking “Malta” and everything else under it to ensure you’re not just matching against sightings at Golden Bay, for example, but rather have selected everywhere the ray might appear.
It’s an interesting experiment you could run, and it would not hurt the HotSpotter algorithm, which only looks for textural correspondence. However, for future ML training (e.g., with the PIE algorithm), we would want to exclude these as we push for full automation. If you do add these as a test, make sure to keep adding the existing, full color photos as well.
Do I need to do anything about the future ML training (pie algorithm)? I’m not sure what this is about.
Final question for now: when we start a match, does it not matter which image we pick i.e. from the binary candidate region or the regular candidate region? or will hotspotter look at all the pictures in an encounters’ gallery ?
Do I need to do anything about the future ML training (pie algorithm)? I’m not sure what this is about.
No. By using MantaMatcher to associate individuals with their photos, you’re building a future data set for PIE machine learning-based training. We look at PIE when we have about 200 individuals with at least three or more photos per individual.
when we start a match, does it not matter which image we pick i.e. from the binary candidate region or the regular candidate region? or will hotspotter look at all the pictures in an encounters’ gallery ?
It looks at the individually selected photo. So run matching on both of them if you want to use the binary candidate method too. One thing to note: HotSpotter internally processes images as grayscale, so going to binary black/white may or may not aid significantly in matching. It is worth the experiment though.
I did think of another question though that I couldn’t find addressed in other topics:
I’m going back to the first bull ray we sighted to start naming individuals (only naming pictures where the pattern is visisble, making the non visible images as ‘unidentifiable’. I am using the names that had been matched by our last matching system and visual inspections and would like to create new ones for ones that had not been assigned a name so far.
My concern is: as I name each encounter, if I do not notice that Br15 and Br75 are actually the same individual - when I try to run a match for Br75 will it reject Br15 as a possible match because its already named or would it still inspect the pattern of an already identified individual?
This answer changes whether I leave individual as unassigned until I am completely sure they are new, or if I name them so that I can start to build an individual ID database.
I just feel like if I leave them all as unassigned the learning process would move too slowly.
Sorry again for all the questions but I just want to avoid mistakes that slowing down processes.
Great question! Adding an ID does NOT interfere with matching. If the match results do show that Br75 and Br15 are the same individual, you can even merge the individuals into one name according to these instructions:
