stars image generation from the nasa skymap star catalog.
libstarid interfaces between numpy, tensorflow, eigen, etc.
identification is the heart of the project.
triangles recurrently sets aside star identities that disagree geometrically, until one remains.
convolutional neural networks are about static patterns.
recurrent neural networks are focused on sequences. it's possible to reduce sensitivity to rotation.
neural machine translation writes sentences about star patterns using a geometrical language, and translates into an identifier language. this can be insensitive to rotation.