It is disconcerting if you are looking at taxon photos - for more info, and among the ‘birds’ is a random tree. It does not provide useful info. Until and unless iNat enables marking this photo as habitat. Or linking 2 obs - this bird eats these berries.
3 Likes
I use an Android phone and have to do the same thing for audio observations.
I use BirdNET because I like the workflow in the app for recording and selecting clips. Oftentime, the bird is calling intermittently, and I can just record until it calls, then select a short clip with the specific call in it, and use that for ID purposes. Even though the android version of iNat can do audio, I don’t like actually using the iNat app to capture media.
BirdNET is also handy because it has an AI that works to identify birds via sound. I’ve always been horrible with things like warblers, so this is super handy for me. BirdNET will also export an obs to iNat (though without the media), so I use that to create a “dummy” observation with no media, but with other important info like the location, date, time, etc. the iNat app doesn’t like the audio format BirdNET uses, so I have to go to my computer later to transfer the audio file to my PC, then upload to the iNat website. I put in the notes for the obs that it’s a placeholder record for an audio file (so that someone doesn’t preemptively mark it “no evidence of organism”) and I make a note of WHICH audio file it is so that I can keep track when doing multiple audio observations. I’ll put a preliminary ID, as well.
The frustrating part is that I didn’t always need to do the audio upload on the computer. I used to be able to attach the audio to the obs in the iNat Android app. But somebody changed something. I’m not sure if BirdNET changed the audio format it saves in, or if iNat started limiting acceptable audio formats via the app. But whoever changed, they broke my previous workflow.
3 Likes
Ah, those are good features! I really should look at getting BirdNet as my skill level for audio identification of birds is not even rudimentary.
Even though the android version of iNat can do audio, I don’t like actually using the iNat app to capture media.
I tried using iOS iNat app to make an audio recording and I also used Voice Memo to record the same bird a moment later. When I compared the 2 recordings, the iNat app recording was far lower quality than the Voice Memo recording (not even useable in that case). And, I could not find a way to edit it (trim/enhance). Both recordings used the same microphone on the same device under the same conditions. Maybe there is a different compression scheme used by iNat for recording audio? Maybe it screens background noise differently?
Anyhoo, after reading your workflow for audio, and knowing the extra steps I go through, it seems sort of convoluted to add audio OBs. But, it sure adds extra dimension.
1 Like
Again, there’s a comment section here you can add as many photos as you want to and making photo not the first one doesn’t mean it won’t be used for cv learning. iNat rules clearly say each pic should contain evidence of organism, so if there’s none - post it in comments.
I’ve been pretty pleased with the quality of recordings from the BirdNET app (another reason I like it). I’ve certainly found limitations with distant birds in noisy environments, but not much to do about that except to try to get closer.
No enhancements, but trimming a piece of audio to include the specific call is dead easy. very helpful if there are multiple organisms making noise at the same time.
I try to get a photo of the organism in question, but usually it isn’t possible as said bird is far too high in a tree (in dense forest) to see (or it’s something like a frog that’s well hidden and camouflaged). I haven’t felt a need to include a photo of the habitat, but if I see something habitat-related that might be relevant, I’ll put it in the notes section.
I do wish there was a way to include video observations. Sometimes the way an animal moves is an important component of identifying it. And maybe it’s moving too much to be able to grab a still, yet video can yield a reasonably good result. to date, I’ve been linking to vid clips (hosted on flickr) in the notes section and trying to use a screen capture as an image. but oftentimes the screen capture is poor quality and while you can tell there’s an organism there, it can be difficult to distinguish.
1 Like
Try adding gifs, it’s not a very well working option, but you may succeed.
“Evidence of Organism” is right in the DQA. If the photo does not contain evidence of the organism that is the subject of the observation, it does not belong there.
1 Like
This applies to the whole observation though, not just one photo. As long as the observation has evidence of the organism, it shouldn’t be DQAed for having one photo that doesn’t.
6 Likes
No, but it does show us that evidence of an organism in the photo does matter.
We keep on getting told that habitat photos are bad for the CV. I really doubt that and would like to see some statistics to back that up. Two main reasons:
-
They shouldn’t matter to the CV: One or two photos of some trees amidst a hundred or a thousand macro photos of a bug or bird should be intergrated out by any good statistical model. They’d be treated as outliers and will just add a tiny bit of background noise to the machine learning’s pattern recognition algorithms.
-
They might even help: We forget that the CV is not trying to identify species the way that we would, by honing in on diagnostic characters of the critter. It’s assigning labels to photos with the right assemblage of pixels that typically get identified as a certain taxon. So if a certain habitat shot is typical for people to actually confidently identify a certain bird (i.e., which pictures become RG), then that habitat might actually help the algorithm get more confidence in the probability peaks it’s assigning to a certain taxon. For example, the oak leaf itself might become part of the signature characteristic of a particlar oak leaf miner insect in the iNat CV model. Or if a certain robber fly tends to eat a certain prey insect, the pixels of the prey might become part of the statistically typical iNat photo for that predator and therefore part of its key signature in the CV model. That’s ok. If the leaf or prey is not actually diagnostic, it’ll eventually get swamped out by other random sightings and we’re back to point #1.
So, I don’t see sufficient cause to actively discourage people from posting habitat photos, which are admittedly useful, even if it’s not preferred.
6 Likes
As one of the top identifiers of audio observations (of birds), I can understand both sides of the argument. Habitat photos can provide useful insight on an observation. In practice, however, I think this is probably more useful for non-avian species. When someone provides a habitat photo in an audio observation, I rarely pay much attention to it. I usually make any inferences I need about habitat based on background species and/or sounds in the recording. I think ideally there should be a way to upload habitat photos, but in a way that it doesn’t impact the CV. As Tristan just stated, I can’t imagine a handful of habitat photos having a measurable impact on the CV unless it’s a seldom observed species. My practice when it comes to my own audio observations is to never use a habitat photo but instead put comments in the description regarding the habitat, species in question, time stamps, individual variation, etc. as I see fit on each observation.
6 Likes
May I ask, how do you record those audios and how do you upload the OBs? (Just looking for an easier way.)
i think the path i would go down to find a solution here is to request a new type of observation field where the value is a photo. if something like that were implemented, that would allow you to attach a habitat photo in an observation field. it might be a way to include spectrograms, too (rather than making them photo evidence of the observation).
short of that, i would personally just capture video of the habitat, load it to YouTube, and then link to the video from the observation. (i think a video is going to provide a better picture of the whole habitat than a photo would.)
2 Likes
And, would that make iNat happy?
After reading comments here, think the downside of attaching a regular ol’ “habitat” photo or spectrogram is the, apparently small, impact it may have on CV.
right. that’s why having a mechanism to record a photo on an observation field rather than as image evidence on the observation could sidestep those issues, i think.
and just for clarification, i don’t think that the main issue with images that don’t directly record the organism or its signs is its impact on computer vision. i think the main issue is that those kinds of images don’t directly record the organism or its signs. a regular person can’t look at a spectrogram and understand what that organism sounds like. a regular person can’t look at a habitat photo sans the organism and see what the organism looks like.
3 Likes
As of lately, I mostly just use the iNaturalist app to record using my phone and submit that way. But I also have a Zoom H4N that I use for high quality recordings, I just haven’t taken the time to go through my massive backlog of recordings from that and uploading them.
yay…more irritating software manipulation. how about not?
? that’s up to you.
3 Likes
Seems like it’d be a good idea in general to have some way to mark a picture as “supplemental information” or “don’t feed this to the AI”.
Also to put the audio observation first rather than having the picture dominate the observation.
1 Like
There are all sorts of things that would be really wonderful additions to iNat. I am absolutely certain that the people who run the site have a very long list of upgrades they would like to do. They all require investment of time and money in a not-for-profit context, by an organization that has experienced enormous, rapid growth in a very short period of time and is dependent on a whole bunch of volunteer moderators and curators to make things work.
Every addition of complexity to the algorithms driving iNat requires up-front investment to write and compile code that doesn’t break the existing functionality. It then adds to processing and data storage loads and the computation cost associated with every observation, forever. That this is an issue is a matter of record in discussions of all sorts of constraints in the system.
What conceivable reason does anybody have to lie about it?
2 Likes