I've been toying around with photogrammetrically matching in-flight bird silhouettes in order to have a plantnetlike bird classifier.
Very much in the data acquisition stage but I'm seeing moderate success in clustering images by bird shape so far. In between other photogrammetry work I'm starting to think about how to use colour information and reading up on how to select and train a ML model for this problem.
I don't know if you've ever photographed birds, but I suspect this would be difficult because shots that are detailed enough to make out enough detail would be tight in, and shots that covered a large patch of sky would not be detailed enough.
Just speculation though, I'm sure with enough resources these are surmountable issues... I'm just not sure what quantity of resources that is.
Yeh this was my initial concern with optical silhouette classification. The altitude and targets are quite small (and rapidly moving). I've thought about using optical flow analysis to track wingbeat frequency etc as I feel this might be a little simpler to acquire.
On wingbeat frequency I'm very curious to play around with mmWave doppler radar (there's some RF on chip stuff around), emitting a relatively isotropic signal might allow for a pretty wide field of view and I imagine you'd get a pretty reasonable wingbeat signal.
