Non-georeferenced photos to train algorithm

Hello:

A technical question.

Can batches of non-georeferenced photographs be uploaded to train the algorithm in a specific area?

Thank you and greetings,
Manuel

what do you mean by non georeferenced? are you planning on loading such photos and then adding location manually to the observation, or are you planning to load these without adding a location at all to the observation?

By area, do you mean geographic area or maybe taxonomic area? The CV model doesn’t take region into account per se (I don’t think) but does note nearness of taxon observation when making recommendations (eg, “Seen Nearby”).

You can see the FAQ on the CV model here: https://www.inaturalist.org/pages/help#computer-vision

4 Likes

Before any of this is answered, most important they have to be your photos. Adding photos from someone else, even open creative commons licensed ones is not permitted under site rules.

If they are your photos, you can add them into casual observations with no location if you wish. With no location they can never achieve research grade and are unlikely to get any supporting identifications.

Please note that even if added there is no guarantee your photos will be used to train the computer vision as once there are a certain number for a specific species the photos used are randomly chosen to ensure a representative mix.

Also the algorithm is not trained based on geography ( not sure if in an area means geography or taxonomy). It is simply trained to try and recognize a species. When used those results are then compared against species recorded in inat observations within the area of the submitted record to tell the user if the suggestions has entered observations nearby

3 Likes

Thanks for your answer. I am a colleague of Manuel, the one that posed the question. Of course we will only upload pictures taken by our team. Maybe I should say we are an Spanish institution (CSIC) promoting the use of iNaturalist all over the Pyrenean range. We are profesional botanists, and we have made a full list of pyrenean plants (more than 4000; most probably not only are listed or observed yet in iNaturalist). In particular we expect a scarcity of pictures for Pyrenean endemic plants. For that reason, it could be very interesting to upload pictures of plants we already know their names, taken in the Pyrenees, even if they are not georeferenced. Do you think that would be possible and recomendable? Our intention is just to improve the matching of the algorithm for Pyrenean plants. Thank you very much

Yes, however there is a number of photos required to be added (I can’t recall the exact number it may be 100) before a species is trained through the computer vision system. There need to be enough photos for the system to ‘learn’ how to identify it.

If these become the 1st photos on the site, it won’t result in them being added to the algorithm.

2 Likes

The CV is trained on verifiable or would-be-verifiable-if-not-captive photos, meaning each observation needs both a date and location to be eligible to be used for training the CV. The date must include the year/month/day, but a general location is ok, like this:

The requirement for adding a taxon to the CV is 100, but that’s not 100 photos, it’s 100 observations.

We also don’t know when the next model will start training, and once it does, it would still be many months before it’s available for use on iNat. So I would say it’s worthwhile to upload your photos – at a minimum other users will be able to learn from them right away – but if your only goal is to improve the CV, it will take quite a while to see benefits.

3 Likes

Hello:

Thank you very much everyone for your answers.

It is very interesting to know how the computerized vision algorithm works, although it is not essential for making observations.

But if our photos are useful to both iNaturalist and the community, that’s great.

Greetings,
Manuel

2 Likes

Thank you very much. Still it seems to be worth to upload a bunch of pictures for people to better recognize them. Particularly for Pyrenean endemic species, for which there might be no pictures at all yet

1 Like