I’m working on a project where I’m testing the use of iNaturalist to assist with species identification in Arctic waters. After uploading and evaluating the predictions (top 3 species suggestions), I’ve realised that many species in this remote area aren’t well-represented in the model, meaning it falls under the long-tail classification problem. https://www.inaturalist.org/projects/pilot-project-assessing-artic-biodiversity
I hope you can help me with a couple of questions:
Does the iNaturalist model use data augmentation techniques like flipping, rotating, or color changes to help improve predictions? Could this help with the long-tail classification problem?
Are there any resources or papers (besides Van Horn, Grant, et al. 2018) that you’d recommend for understanding how to handle this kind of issue?
Would love to hear your thoughts and suggestions! Thank you
/ Annika
If I’m not mistaken, a species can only be suggested by the CV if it is part of the CV training set and for that, it needs at least 100 RG photos (photos and not observations) to enter the CV algorithm. To get a wider variety of suggestions in your area, it will help if you add observations or identifications to iNat so that more of your species pass the threshold and will be added in the next CV update. See also a discussion here https://forum.inaturalist.org/t/how-are-photos-selected-for-cv-training/42403/8
Yes, to clarify my short reply above, iNat does flip, rotate, etc., photos used for training, but there still needs to be a minimum number of photos available on iNat before a taxon is added to the model, as @frousseu and @dianastuder mentioned.
