Abstract: A camera tracking system for computer assisted navigation during surgery. Operations identify locations of markers of a reference array in images obtained from tracking cameras imaging a real device. Operations determine measured coordinate locations of a feature of a real device in the images based on the identified locations of the markers and based on a relative location relationship between the markers and the feature. Operations process a region of interest in the images identified based on the measured coordinate locations through a neural network configured to output a prediction of coordinate locations of the feature in the images. The neural network has been trained based on training images containing the feature of a computer model rendered at known coordinate locations. Operations track pose of the feature of the real device in 3D space based on the prediction of coordinate locations of the feature of the real device in the images.