I have a collection of camera trap images featuring an array of interesting wildlife.
I have used an AI algorithm to automatically detect and mark images where there are animals present (but not classify the animal).
I now have photos containing an image of the animal, with a box drawn around the animal to show where it is. Here is a similar example I found online (Image credit University of Minnesota, from the Orinoquía Camera Traps data set.)
The question is whether these images are suitable for upload to inaturalist?
Im preety sure these are okay to upload, aslong as the picture itself of the animal isnt ai generated, and you put the location/date and other info correctly.
Good points, thanks for your advice. We will still check and upload each detection manually, but it should save lots of time searching through the empty shots. Thanks!
i’m not sure if you would want to do as tight a crop as an AI detection bounding box would be. it’s not necessarily the common view, but i would prefer an uncropped image, even with lines drawn in, to one that completely crops out everything but the subject.
There is in fact an option to separate the full framed images without putting a bounding box at all on future processing. After hearing the community views, it may well be worth me reprocessing previous ones too and then uploading them without the boxes here, since we will need a human verifying each upload anyway.
Appreciate everyone’s views on this, some great insights and seems like it may become increasingly used in the future.
Hm … I honestly do not really see the point for this – any human would immediately see the cat in your example without that bounding box? There are harder examples for sure where the animal is very small in the pic, but I think that in these cases a crop is preferable in any case?
AI can be useful, but it can always hallucinate and has an immense energy footprint; I personally would not use it where I do not see a substantial benefit.
This is software that you run on your computer to scan hundreds or thousands of game camera images to find those with animals present in the scene. I think it’s sorta like the iNat computer vision in a way–except it determines if an animal is in a scene. https://github.com/agentmorris/MegaDetector
It’s not like Large Language Models which are creating sentences (generative AI, e.g ChatGPT). The immense energy footprint associated with generative AI doesn’t apply here.
Thanks. But I am still struggling to see the purpose (note that I do not say that you shouldn’t do it; I just try to understand why you think this is helpful). You do it to help identifiers to see the animal? Do you have an example of a case where this is useful? People sometimes draw circles around the animal – yes, but they usually do it when there are multiple species seen in the pic, and the AI is of no help in such cases since it does not know your intention, right?
Hi,
As an example, say we have a network of over 10 camera traps, which sometimes take over 1000 photos on a weekly/monthly basis. Checking through all of these images manually for animal presence is quite unfeasible for a small team and as such prevents the team from sharing any observations. Software such as this is therefore deployed as it automatically searches through these thousands of images, and detects where an animal is present. It can draw a box around the animal, and then transfer it to a separate folder, so that humans spend their time looking through only the images which have something in, rather than the thousands of empty shots triggered by things such as the wind. It isn’t perfect by all means, but in the long run enables more data and observations to be shared with the community!
Ok I see. So you have camera trap shots that already have these bounding boxes anyways and are wondering if you could upload them (of course you can). There was no intention to add those boxes just to help identifiers see the animal?
If you have many of those camera trap shots: I think it would be not too difficult to write a script (no AI needed here) that crops the image to the bounding box while keeping some extra margin (so to not get a tight crop).
Yes, exactly correct! The bounding boxes were purely a result of using this software to help remove the empty shots from the files. As you point out, once a human is checking them for upload here, they add minimal benefits in this case.
Interesting idea about the extra margin. The tool in question allows separating the whole images without a box at all, so for future identifications we will probably just look to use this method I think.