In a project that I manage, participants sometimes upload observations using a collage (for example: overview, detail, habitat etc. combined into one image), instead of uploading multiple individual photos. Here are two examples:
Do such collages interfere with or confuse the iNaturalist Computer Vision model? Could they negatively affect species identification, or is it generally fine as long as the organism is clearly visible?
Collage would be fine as an extra. But taxon specialists would prefer single images so they can zoom in on relevant details? Your second example is only the collage.
Theoretically yes, the CV is trained by the data input in. But the amount of impact depends on how much data its trained on. So in small amounts its not an issue, but it could potentially create issues if it is in larger amounts on the same taxon.
The CV can cope with this. It’s we identifiers who have a problem. With the collages, we can’t zoom in on details we need to see. So I’d say, please discourage this (or encourage posting a collage in addition to other photos) but not for the reason you suggest.
My question is why do people bother with the photo collage in the first place? inat isn’t an art project. why bother with the extra step of making the photo collage when inat can handle multiple photo uploads just fine? maybe people are creating these for their social media channels and then just using that for inat out of laziness?
but I agree with other folks - if they’re going to include it, it should be in addition to unaltered, original photos. I sometimes include edited images as extras to assist with ID. sometimes cropped images to highlight something that’s camouflaged or even circled if that camouflage is particularly cryptic. text on images is unnecessary, given there is a description field.
That’s curious - I can zoom in on the details without any problems. Are you viewing the photos in a different way to me? I’m using a desktop browser. Whenever I need to zoom in even further than the iNat viewer allows, I just open the photo in a different tab. I find I need to do this almost all the time anyway, since the large majority of photos are uncropped. Some people include a cropped photo in addition to the original, but I never find this particularly useful. And I don’t see how stitching them together as a montage would necessarily make much difference.
I think you mean montage, as a collage creates a completely new image by merging together several disparate parts. But either way, the Computer Vision shouldn’t be affected by the juxtaposition per se as it has no … well, “vision”, as such. It just looks for statistical patterns in various aspects of the raw data, so it has no way to assess the image as a whole. A montage of several views of the same subject need not “look” much different to a group of individuals of the same species. (Unless the latter included e.g. different sexes or life stages).
A more serious potential problem with your particular examples is the framing separating the sub-images. That creates a distinct, consistent feature that could be picked out as statistically significant (assuming enough similar examples were included in the relevant parts of the training set, that is).
In some of my earlier observations, I did not know that an observation can comprise more than one photo. I made collages thinking that it was necessary if I wanted all the different angles in one observation. I suspect that the problem @DianaStuder has complained about – different features of the same organism spread across different observations – comes from the same lack of knowledge.
iNat downsizes photos to a maximum of 2048 x 2048 pixels. So if a user makes a montage out of 4 images, the amount of detail you can zoom into is often considerably less than if each of the images had been posted individually. This can matter quite a bit for observations of organisms where tiny details may be crucial for ID.
There’s a pixel dimension limit on iNat. In my experience, that pixel dimension limit already causes loss of diagnostically important data when uploading separate images. Combining multiple images into one just exacerbates that issue. It’s basically the equivalent of uploading separate tiny images—which is also something people do sometimes, unfortunately.
Anyways, I don’t know about the CV but I agree with others that the collages are dramatically less useful for me.
There is another aspect: iNat restricts and reduces the resolution of every single picture. To save the maximum of details for an observation it’s best to cut out what you want to identify rather than combining pictures.
One other potential problem is illustrated by the second example (mayflies). In this montage, there are pictures of different individuals, with different numbers of tail appendages. Because it is a single image with potentially multiple taxa, there is no way to click “No” on “Evidence related to a single subject.”
This need not be true at all. If you take quadrants of four 2048 x 2048 photos with the same resolution and assemble them into a montage, it will have no effect whatsoever on the final resolution. There are obviously countless bad ways to prepare photos of any kind, so that’s not a valid argument against any particular format. It’s not difficult to create perfectly valid montages that don’t negatively impact their usefulness in any way.
Having said that, this is all a bit moot, since I don’t think anyone is actually arguing strongly for the increased use of montages in general. However, there are a few legimate narrower use-cases (such as taxon photos, as already mentioned). Another one stems from the number of images needed by some observations. iNaturalist doesn’t really set a limit on that number, but if it did, that could potentially necessitate the use of montages. Some taxa require a lot of detailed photos to key out in full, so it’s not implausible that anywhere from 10 to 30 photos could be needed to do it rigorously. I regularly upload observations to another site which has a hard limit of just four 1600 x 1600 photos per observation, so montages are often an absolute necessity there. I have never once received any complaint from the validators for including such photos (quite the opposite, in fact). At present, the site in question automatically imports RG observations from iNaturalist (via GBIF), so that provides a legimate reason for some people to make use of montages.
As an identifier, I don’t really understand some of the push-back on this. It’s up to the observer to decide how best to prepare their photos. If the only evidence they provide is a photo-montage, it would feel rather petty to complain about it. There are literally millions of real examples on iNaturalist that provide much, much worse evidence than that. I really wish even a small fraction of the extra effort that goes into creating a montage could be spent on each one of those.
Sure. But in practice I have found that this is not the case for most montages I encounter. A lot of the time I find myself wishing that the observer had included the individual images in addition to the montage.
I have not seen any complaining here. It is, however, legitimate to have preferences and to comment on what sorts of practices work better or worse for us as identifiers.
And from the perspective of an observer, if I am putting effort into preparing my photos before uploading them, it is likely that I am doing so because I care about the quality of my observations and how they are received by others. If it turns out that what I am doing is not actually useful or even counterproductive for others, I would certainly like to know about it so that I can decide whether I want to continue to devote my energy to this or whether it might make sense to modify my practices.
Actually I’d prefer that a montage not be the default taxon photo, because taxon photos are used as thumbnails when selecting a taxon as an ID and I find it useful if they at least provide a sense of what kind of organism I have selected, as a sort of visual warning to help recognize wrongly typed IDs or accidental selections from the drop-down menu. This requires that the subject be large and contrasty enough to recognize its gestalt, which is unlikely to be the case with a thumbnail of a montage image
I do not understand your point. The montage will be downsized to the final 2048 px resolution so if the original photos had 2048 each, they will be 1024 now.
At the risk of labouring a point, commenting that something is worse than desired expresses dissatisfaction, which is precisely what a complaint is.
There’s no need to resize the final image at all.
Take any 2048x2048 image and divide it into four equal 1024x1024 squares. Now reassemble the pieces in a different configuration, and you have a montage with exactly the same resolution/dimensions as the original. This process can obviously be done with four pieces from different images of the same size.
I’ve got to correct myself here. It’s many years since I last tested this, and I’ve now found that there’s no longer any obvious limit to the number of photos that can be imported from iNaturalist - even though the old restrictions still apply when uploading through the site’s own interface. This is achieved by the simple expedient of linking directly to the iNaturalist Open Data on AWS.
So, apart from taxon photos (where the guidelines suggest using as few photos as possible), it’s hard to see why it would ever be necessary to use montages on iNaturalist. Nevertheless, I’ve yet to see a concrete reason to actively discourage anyone from using them, and I still feel it should always be left to the observation owner to decide how best to present their photos.