Yes, but you needed to edit the observation to make the intended photo the default observation, then request CV suggestions. In iNat Next, you can choose a photo on the CV suggestions screen without having to go back and forth and edit the observation’s photo order. For example, I’ve chosen CV to analyze photo 2 in this observation:
7 Likes
I will clarify that I was talking about the android application, I don’t know about the ios app
2 Likes
I use Google Lens for tricky obs - there I can crop to - that bird - top right at the berries.
1 Like
What’s iNat Next? I saw a couple posts about it, but it’s still in beta testing right? Is it for all versions and platforms, or just for iPhone?
1 Like
To me the main thing for AI identification is for the original poster, because you want any further IDs to be made by people who know themselves how to ID, for which the AI help is more a way of saving typing, being reminded of the exact name, or jogging the mind in some way. As a result I would focus any AI improvements around the posting experience.
Now I can understand not wanting to analyse 20 pics on an observation, I would think a couple would be achievable (covering most obs), by default the first n photos - presumably something analogous to whatever plantnet does with its 1-4 pictures to create a single computed score. For those wanting more granularity a box could be provided on the left side under the advanced fields where picture indexes can be entered (blank meaning all) with a button to provide a report of recommendations and scores for each picture so the user can decide how to combine them (or just a button which pressed prompts for the indexes and then creates the report). I’m sure the number of people using this feature would not pose an extra burden.
For people who post random pictures in 1 observation, I think when IDs on the picture set are fairly incongruent, they could be asked to confirm if it’s all one plant, and if not to automatically split it into separate observations for them. A flag could be set for observations that are submitted as incongruent to help find and sort them out. As well as incongruent IDs, incongruent picture dates or embedded GPS can be used to issue the same prompt. Perhaps curators could also have a button to split observations into several to help tackle the problem, perhaps as a list of comma separated picture indexes, with semicolons representing observation breaks eg 1,2,4;3,5
I think the CV should evaluate an observation based on all the photos. That’s what we do in the field when we see something, right? Things I find and try for good multiple photos are: mushrooms, dragonflies and birds. These can all require views from multiple angles to determine an ID.
3 Likes
Yes, I noticed the new app lets you get an AI identification separately for each photo in an observation, which is a big improvement. It was one of the main improvement that got me to switch to the new app. (Another was that the new app provides a numerical confidence score now.) It’s still not really what I was originally asking for, though.
With plants, at least, the combined information from the different parts of the plant (top and bottom of leaves, structure of twigs, bark, upper and lower sections of a plant stem, top and bottom views of a flower, etc.) can let you make an ID that’s much better than just one part (like, a leaf, or the bark). It’d be nice if iNat could do that for you, and merge the information from the various photos intelligently.
2 Likes
I apologize for reviving this very old thread, but I recently filled out a survey for identifiers and realized I forgot to mention a major roadblock I consistently encounter, which seems related to the topic discussed here.
I primarily use the web version of iNaturalist, and it still appears that the computer vision model only provides its suggestions based on the first photo of an observation. It remains frustratingly impossible to select a different photo within that same observation to get a CV suggestion based on a better angle or feature.
I wanted to check if I have missed something recently or if this limitation is still in place for the web interface. If this has been addressed in a more recent forum thread or update that I missed, I sincerely apologize for bringing up an old topic!
13 Likes
17 posts were split to a new topic: Computer Vision and Text Entry when adding IDs