Different observations for different sexes


I have a question regarding the upload of different sexes of a species. I typically upload both sexes of a species into one observation, but I have noticed that one cannot assign a sex to each photograph in an observation to distinguish them. I am not sure if the AI uses this information for anything, but if it does, how would it know which ones are male/female if they are in the same observation? Is the idea to upload them as separate observations?


1 Like

Will Computer Vision in the future recognise annotations like ‘sex’ and ‘life stage’?
If so, what is needed to make it happen?

Doubtful, the difference in sex is based on coloration in some vertebrates, but by morphology in flies, so the CV would have to have it specific for each taxon.

Idea is each specimen should be in separate observation, for now cv doesn’t learn on annotations.

1 Like

Yes, pairs of male and female would indeed be two observations, one for the male and one for the female. If they are both in the same image, you could use the “duplicate” feature for the observation and then add notes in the descriptions in addition to annotating the sex on each to clarify that one is for the male and the other for the female. (Without those notes/annotations, it may get flagged as a straight-up duplicate.) Doing it this way has the advantage that both observations can be accessed by clicking the info button on the picture. If you upload them separately, that will not be the case but you could link them together e.g. using the “similar observation set” observation field or just by adding links to the other observation in the description.


For pairs of waterfowl, I usually use one record to cover both individuals and make a note that it’s a pair. On occasion the individual birds are in separate photos in the same record if I can’t capture them in same frame.

One observation is for one individual, never for two. You need to make two observations one for each sex. It could be the same picture with notes describing the individual for which the observation is made.

For species with a large amount of sexual dimorphism I think the current AI pretty much has to just be learning to ID the two appearances separately and then just learning that the same label applies to both appearances. This is also how likely how a human would have to learn it.

It would be interesting as a completely separate step though, I’m sure there are plenty of cases where you could learn to recognize features of sex without actually being able to identify to species. Similarly for plants it seems like it would be pretty easy to learn to annotate ‘flowering’ without being able to confidently ID even to class.


Maybe, but it’d be expensive to run the training with 3x as many classes (or whatever), and problematic for clades that are sometimes but not consistently annotated. It would be much cheaper to run training with just a few possible labels but ignoring the species, to the point that it might be practical for a third party to do with just the API. Existing publicly available semantic labelling networks could probably already do ‘flowering’ off the shelf with no re-training at all.

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.