Hi Georg, team,
Reporting back with some awesome progress - we've finished the first run with 10% training data from Stellarium on grayscale Resnet 18 neural network and got
67% success rate with a star identification! Unfortunately I can say how many epochs it took to get there as the training script died (leaving trained network weights settings in a file), but still I think this is super-awesome - demonstrating that this concept works! (One training epoch on 10% data took ~28 minutes.)
Next stages are:
1/ Rerun this with Resnet30 + on a whole training set (pictures needed to be scaled down to 224 x 224 grayscale). Hoping to get a success rate above 90%!
2/ Start working on recalculation back to celestial coordinates
3/ Prepare Stellarium Proof-Of-Concept
This is where I've been thinking if you can share your thoughts. My current rough POC plan is to:
1/ Ask Stellarium to pick random datetime + location
2/ Identify 4 brightest stars above horizon
3/ Return their screenshots + inclinations + datetime
4/ Log everything in a file - including data for validation (star names, location)
5/ Identify stars with pre-trained network
6/ Use inclinations + datetime to identify location
7/ Compare with validation data from Stellarium
I think this should be all ok-ish, the only thing I am not sure about would be point 2/. Would you be able to point me in a direction if there are any analytics available in Stellarium to be able to identify 4 brightest stars above the horizon, or suggest any clever trick to get around this?