Utilize state of the art deep-learning techniques to detect a satellite’s model and its relative orientation using images captured of the satellite.
We optimize the hyperparameters of four CNNs from the ImageNet challenge in a fine-tunning setting.
We utilize Unity Engine and NASA's 3D Model Library to generate images of satellites from various perspectives.
We utilize kolomogorv filters to distort images in our synthetic dataset to mimic images of espionage satellites taken by earth-based astrophotographers.
We utilizes Amazon Web Services to train our models in parallel.