Please do post photos of these shells, showing each species’ range of variation and and geographic distribution. Those photos will be valuable for anybody working with the group!
2 Likes
Will do!!!
in a few days…
I’ll be on so many leaderboards…
unless my schoolwork gets in the way 
4 Likes
Also in crayfish asymmetry is frequent. In some species, the left gonopod of the male is shifted posteriorly. In most females the sperm receptacle is asymmetrical and in some species, the configuration is randomly flipped, but never flipped in others. I suggest that flipping should probably be taken out of the training process.
I would suggest that you photograph a few more of each species, making some additional observations, because micros are so poorly represented here on iNat in general.
2 Likes
Please please please post as many as you like! It will help iNat greatly!
2 Likes
How often is the asymmetry the main diagnostic for the species? If there are a small number of available photos in the CV training set for a species where the true configuration is randomly flipped, just by statistical chance it is possible the dataset could appear rarely/never flipped, causing the CV to erroneously learn a spurious asymmetry feature. I presume the augmentation would stop doing the flipping once there are enough photos it is no longer needed. So in that case the flipping might act similar to label smoothing (where it assumes there is, say, a 10% chance the ID it is training on is incorrect) in that it can try to prevent the CV from being overconfident about the significance of a feature based on a smaller dataset.
For the Pending species - you can check how many obs are here. And look for more to add. Either your own (make sure you include enough info for a good ID). Or plough thru Needs ID, they may be waiting in the queue. One of the benefits of picking thru Pre-Mavericks is tipping patient biodiversity into CV. Particularly if we have acquired a shiny new taxon specialist for a neglected corner!
We get a substantial chunk of new to CV each month. Latest update is due?
I appreciate this whole topic, and it’s been very interesting to read. It’s also the perfect thread to ask a question myself that I’ve been wonderinf about for a while regarding the CV model. I understand the ~100 obs including verifiable-if-not-captive, and am just wondering about that. For a few species in the latest update are almost soley on iNat from captive observations (eg Pygmy Hippo and Fishing Cat). Is there any measures in place to make sure that the model does not learn to associate that species with a ‘zoo enclosure’ look, eg looking down on the animal at a certain angle or fences or bars or comparative lack of vegetation? Or has that just proven not to be an issue?
And similar thoughts about animals that are largely known from camera trap imagery, could the CV model learn to associate a certain species with the look and text of camera trap images and eg suggest any camera trap image of an ungulate as, say, a moose or a white-tailed deer where there are thousands of camera trap images used as input for the CV model?
Of course, these aren’t major issues, afterall the CV is just meant to assist and guide, it’s not building RG observations by itself, but I’ve just been curious if these are things you’ve considered and perhaps have built protections against into the model.
Thanks,
Nathan.
6 Likes
And not only the CV, but humans too can learn from looking at more images of shells of small marine mollusk species.
6 Likes
I have added about 15 observations each for the flat cardita ( Ptromeris perplana) and for the Grey Pygmy Venus (Chionerynx grus). Each observation is of one shell with four pictures ( front and back on two backgrounds) which is about 50 pictures for each species. I hope to post some more species but this is what I have done so far.
Here are the photos for the cardita https://www.inaturalist.org/taxa/1027583-Pteromeris-perplana/browse_photos
And here for the Venus
https://www.inaturalist.org/taxa/461901-Chioneryx-grus/browse_photos
Please note that you have to sort them by “date added” to see mine
I have tried to vary size, coloration, angle, and background. Because I was trying to represent all of the color variations, a lot of them are beach worn.
5 Likes
Nice! Note that the date and location of the observation should be those of when and where the shell was found, not the date of upload.
3 Likes
The location is accurate. I haven’t bothered with the date cause I’m not sure that matters very much with seashells. If it really bothers people I can go fix it, it’s just that it is a lot more work.
1 Like
I understand. I agree that location is more important than date in this situation. Perhaps you could bulk-add a general comment that the date of collection was between such and such a date, so that the information is there if anyone wants to query it.
4 Likes
Yes, the date should be accurate if at all possible. We don’t know what data will be important to researchers for answering any given question, but the expectation is that the data is accurate. In many cases, inaccurate data is worse than no data at all.
6 Likes
I will go fix them then
4 Likes
Thanks!
2 Likes
Yes, and valuable to anyone trying to learn to ID the species.
2 Likes
Also the date being accurate affects ‘seen nearby’ I think?
I just scrolled through some Pygmy hippo observations and I think what you are concerned about actually is happening; it is almost always first-suggestion correct on the pygmy hippo observations that look like they are in captivity, but it does think the handful of observations that do not obviously look like a zoo:
https://www.inaturalist.org/observations/101712072
https://www.inaturalist.org/observations/844282
are first-suggestion common hippos. However, it is right on this wild observation that happens to look like the captive observations:
https://www.inaturalist.org/observations/138111554
I think this is unavoidable because ‘wild hippos aren’t pygmy hippos’ is unfortunately a statistically solid rule. This could be fixed with more wild observations, but probably people should not be encouraged to harass critically endangered species for photos.
2 Likes
I am such an idiot! After just reading about sinistral vs dextral shells (around post 30 in this thread) I IDed one of my own shells as a left-handed species when the shell itself was right-handed!
https://www.inaturalist.org/observations/163852273#activity_identification_1db7fba8-2163-4a24-8140-160f35dedb97
by the way, I added ~50 more photos each for the three tooth carditid and lunate crassinella
3 Likes