by Hooray » Sat Aug 26, 2017 6:36 pm
You may want to tell us in more depth what you are trying to do, I am a little familiar with ML, but I don't quite see the use-case you have in mind so far ?
Speaking in general, the nature of a ANN is/works such that it does not even matter if you -the programmer- know anything about the data that you provide, as long as normalize it such that it works in the form of doubles, you can literally pass in all sorts of properties from /orientation, /position, /sim/current-view, /sim/view and let it determine what is relevant/useful to come up with the kind of algo/heuristics that you want it to "learn" (back-propagate to).
Besides, assuming that you are looking for a way to compute/render a bounding box for the aircraft to crop the size of the image so that your ANN/CNN has less work to do, you may be better off using some C++/OSG code to render such a bounding box.
Alternatively, if you really don't want to touch any of FlightGear's C++ code, and want things to do the ML-only way, my suggestion would be to use a layered ANN, with one of the first layers trained to recognize airplanes/aircraft from different perspectives, and return the image coordinates accordingly - so that you can use the first two ANNs to extract the sub-image and/or render a bounding box.
If you literally need coordinates/dimensions in meters, I guess one thing that may work is parsing the FDM itself and annotating the image using the fdm sub-tree in the property tree - in addition you could dynamically create a livery that contains information/markers to extract dimensions from it.
As you can see, there are probably several ways to approach this, but what is ultimately feasible depends really on the concrete use-case you have in mind, and the tools and expertise available - e.g. a hobbyst will usually have very different constraints compared to a PhD student in machine learning ...