See this observation: https://www.inaturalist.org/observations/198167559
This crayfish apparently detached its chelipeds (arms), then for photography, the photographer tried to reattach them, but screwed up swapping the left and right ones, making them appear upside down. It probably doesn’t matter too much, but such specimens shouldn’t be included in the AI training dataset. They should be annotated somehow to prevent this I think. Any ideas on how to annotate this condition?
i don’t think there’s a way to prevent this, and i don’t think one photo is going to throw off the training.
i would just make a comment on the observation, in case other folks come across it.
Each image used to train the model is rotated and flipped a few times and the model is trained on those variations so that we can get more use out of each photo. So this shouldn’t negatively affect the model. And yes, there’s no way to annotate it, plus annotations are not taken into account when choosing photos to include in the training set.