Sharpening loses information but gains edge sharpness in images which makes them nicer for humans to look at but does it help or hinder the iNaturalist Artificial Intelligence identification process?
From personal experience (NZ spiders), it couldn’t make it any worse. Perhaps it depends on the type of observation though.
Yeah, AI has issues at the best of times but if you have a taxonomic group that is readily identifiable from photos (e.g. moths) and has a lot of correctly identified photos in the system for the AI to learn from then it is very good IMO. Here in the UK I get very good hits each time I try an ID on a new moth. As I processed my photos today though I just wondered if I should be sharpening them … they look better to me but AI is one of those mysterious things that few understand and it might be that it’s best not to adjust the images and lose data. But then you can argue that if the system learned from sharpened photos (because we all do it) then it might be better to sharpen them.
It shouldn’t matter - the training data will contain a mix of sharpened and non-sharpened photos already.
I’ve noticed uploading very high res close-ups can throw the autosuggest because in some taxa its trained mainly on lower res more zoomed out images. But this is just all the more reason to add more variety into the mix to better the training data in future.
Welcome to the forum!
I can’t really say too much about this. I never use Sharpening, but then I have very poor photo processing software. I only change brightness or contrast. I have this belief that the image should look as much like the original as possible. I don’t know why I have this belief, but there it is! I also don’t know too much about how the AI works.
I have identified thousands of moths, and many of the images are dodgy (but identifiable). I do know of a person who was running a test to see if AI or human ID was better, but I don’t know if he has compiled the results yet.
I don’t know if it matters for the AI, but I default to not sharpening as I also think the image should be more like the original if possible. If a user wants to sharpen for some specific reason to allow them to ID, they can always download and do that.
FWIW, I tend to use sharpening / denoising for birds and mammals but not for insects (particularly leps and odonates) or at least I set it a lot softer, because I find that it often kinda messes up some of the smallest details in the wing veins / patterns and doesn’t seem to be helping that much either. YMMV with the results of course, I guess if you had a very close up macro shot it wouldn’t be a problem for example.
It’s not clear to me whether you’re asking if sharpened photos will have a negative affect for training the Computer Vision model (we try to not call it AI, for what it’s worth), or whether posting a sharpened photo will result in a less accurate ID suggestions from iNat’s Computer Vision?
I can’t speak much on the latter, but I think it would depend on how much sharpening you add.
As for training the model…I don’t think it will have much effect. The model is trained on lots of different images which will have varying levels of sharpness. The model is trained to recognize iNaturalist photos of organisms, not organisms themselves. So the model reflects the photos that the community posts.
Seems like minor sharpening shouldn’t have any impact. Of course you can really distort a fuzzy image with heavy sharpening which might not help anyone including the CV. At the opposite end, a heavily pixelated image such as from a smartphone might be hard for human or machine to interpret.
Thanks tiwane - I think you and a few others answered the question actually by the logic that the images that the CVM uses are all with varying levels of sharpness so it wouldn’t matter what level of sharpness my photos were when I uploaded them for ID, because there is already enough variation in the learning set to cover all eventualities :)
Do you actually have stats that show how the number of training images on iNat relates to the degree of success in making new IDs? How many photos does it take for the CVM to have a high rate of success?
FWIW, I edit all my images, which includes sharpening, prior to uploading to iNaturalist. I have not had any issues with the CV getting me to the species, or very close to it. The only “issue” I see is, at times, I get an incorrect species (in the case of some organisms, e.g. spiders, there are clearly mis-identified images that the CV is now using as THE identification).
Someone commented on earlier in terms of reduced clarity in wing venation in I think, odonates. I get the opposite result. I also edit with “contrast” and so any such detail is enhanced in my images, not reduced.
I’ve ended up with some pretty mediocre images that, after some editing, CV has worked wonderfully. Perhaps it would have worked with the original but I only have so much time and so I only upload edited images.
Two years old now, but in this blog post, it says 100 photos minimum for a species/node to be included…and up to 1000 photos are used where possible (…so the more the merrier up til that point).
But in terms of actual success, in inverts it really depends on how much training data there is for species in the rest of the genus/family (or other confusion species) as well. Regardless of how many photos are included, if a confusion species is not included at all in the training data then the model will not be able to discern that alternatives even exist, so will be overconfident in its suggestion.
This seems to have been tempered a little in the last model update though - I find the autosuggest mainly gives genus-level suggestions, which is encouraging.
I’ve wondered the same, but not so much for regular sharpening, but for neural-networked (so-called ‘AI’) sharpening now available. I use Topaz AI sharpening as well as their upscaling product (Gigapixel) and denoising app (as well as other NN-based software). Each process makes a lot of highly ‘educated’ guesses based on additional databases.
When you get too many pieces external to iNat’s CVM system making digital guesses, can that affect the system’s success?
This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.