Images not getting matches

What Wildbook are you working in?
Grouper Spotter

What is the entire URL out of the browser, exactly where the error occurred?

Can you describe what the issue is you’re experiencing?
A lot of images are returning no matches at all. This seems to happen very often even if the fish has very defined pattern that could presumably look like other fish and then some images get “matched” with many images most of which are not matches. Is there a reason behind this?
In the first url example Grouper Spotter - Citizen science and facial recognition for grouper this fish was provide a couple images to match (none were matches) but not with the image that it was previously matched (an identified) from a different encounter Grouper Spotter - Citizen science and facial recognition for grouper | Login. Why isn’t the platform picking up this match?

Can you provide steps on how to reproduce what you’re experiencing?

If this is a bulk import report, send the spreadsheet to services@wildme.org with the email subject line matching your bug report

Hi @accfish

Thanks for reaching out. I can see based on those links above that HotSpotter compared the presented pattern to over 2500 other patterns.

This is a good indicator that HotSpotter is considering other matches.

HotSpotter presenting no or few false positives is also a desirable state. Most of the time in large image catalogs the answer is “no, this is not a match.”, and we want HotSpotter good at throwing those out. HotSpotter will present up to 12 potential matches in this default view, but if there are fewer than twelve with scores >0 (i.e. no evidence of matched patterning) , it will only present those or none.

In the first url example Grouper Spotter - Citizen science and facial recognition for grouper this fish was provide a couple images to match (none were matches) but not with the image that it was previously matched (an identified) from a different encounter Grouper Spotter - Citizen science and facial recognition for grouper | Login. Why isn’t the platform picking up this match?

Of the two annotations for LC000238, one was labeled as viewpoint “up” while the other was “left”, as such HotSpotter would not attempt to match them.

Thank you,
Jason

1 Like

For a Sighting like this for some reason the platform has boxed the fish twice once as left and then once as up and matched it with different images
Does that seem odd?
I’m not sure how to deal with this situation because the match of the smaller box is good but I’d like to keep the larger box as it shows the whole fish but the larger box is not matching with the 257 Individual because it is annotated as up. Should I reannotate the whole image and rematch?

Is it better to have both an up and a left annotation in cases like these?

On a similar topic do you think red lasers on the fish (or lack of when trying to match) could be impacting matching success? We have about 1600 image in the system now with lasers (keyword laser) and more often than not the image either returns no match or the match returned is an image with a laser on it (not always). Do you think this is an issue? If so is there anyway to get around it?

This one has and example with a lot of match attempts most with laser a couple without.

It does happen sometimes with machine learning detectors. If it happens in a large volume then we need to look at how our ML models are tuned (i.e. they are finding too many false positives), but in small amounts thes eare expected.

One thing to be aware of is that if Wildbook finds two annotations (as proxies therefore for two animals) in one photo, which breaks the definition of an Encounter (one animal at a point in time), it will create a second encounter and send each annotation past the first one to its own Encounter. So one of these annotations is held on this Encounter:

https://www.grouperspotter.org/encounters/encounter.jsp?number=12921854-4861-4d96-8469-6b7a8b4a41ec

and the the other is held here:

The best is to delete the Encounter for the annotation that is the least accurate, which would probably be this one:

https://www.grouperspotter.org/encounters/encounter.jsp?number=12921854-4861-4d96-8469-6b7a8b4a41ec

Unknown. It does add uniformity between images (i.e. the red dots), and they are uniform in relation to each other, so HotSpotter may very well start scoring a potential match between two such images higher because there is now a small amount of similar texture (the dots). But these fish do have a lot of other texture, so I think it will not likely change match results significantly.

Anecdotally, from this match result I just ran:

It looks like the dots are pretty well ignored, which is good.

Thanks,
Jason

It is great to have as many images of the fish as possible, but my understanding was that the project was going for left and right images showing the natural pattern prominently. If you expand your study design to include top images, then these may be matchable.

Just FYI: left and right side images are matched to corresponding viewpoints in a matrix of surrounding viewpoints. So a “left” will be matched to: topleft, bottomleft, back left, and frontleft. But a left is too far away from a top, front, back, or bottom to be matchable at such as extreme angle.

Thanks,
Jason