This is the CV model as I understand it
- Select clades
- Select training set for each clade and train
- Keep or reject based on testing
When a photo is processed by the CV model, it returns visual similarity as a number against each clade it was trained on. The top ten or so visually similar clades are passed on to be processed as suggestions.
This request is a change to clade selection.
Your proposal neither supports nor rejects this request so it does not belong on this thread.
I think that your proposal is great, with some clarifications. It needs a separate Feature Request as it applies to the geomodel and the suggestion logic not the CV model.
The two proposals are not alternatives. For a correct suggestion, the CV needs to return a correct clade at taxa, genus or higher level and the geomodel based filtering needs to accept it at the same or higher level.
Yeah I initially made a separate feature request but it was rejected for being too similar to this one.
Since no one had mentioned the geomodel yet I wanted to make everyone aware that the CV and geomodel are closely linked, clades only have a CV if they have a geomodel and vice versa. Also, there is a possibility to solve this issue in some cases by adjusting the geomodel rather than the CV.
Ultimately though, I want this issue fixed so just trying to keep the conversation going and encourage people to come up with creative solutions.
1 Like
The scenario you described is valid and I havenât seen it discussed recently.
I would like to see your request.
1 Like
I would expect that the underlying architecture would need to change into something more akin to an LLM/MML. The AI would need to be trained with various taxon descriptions for best results instead of just photos and metadata and whatever it does now. Done right, it could suggest never before photographed species and say why it thinks that. Done wrong and itâs a massive bloat that takes too long and doesnât do much.