Model description and training experiments:
Our approach is to use pre-trained deep learning models as generic feature extractors. This has been shown to be effective for transferring low level representations of images learnt from an image dataset of multiple millions of images (Imagenet) to solve problems using deep neural networks in multiple application domains (Transfer Learning) \citealt{Yosinski2014,Saad2008,Bengio2012,Kumar2016,Shao2015}. The core of transfer learning approach is to train neural networks by adapting either parameters or features from a strong classifier trained on a data rich task to a new data deficient task. We use a much smaller network for doppler angle estimation by using these features, which also implies a device agnostic modeling process. A schematic of the complete flow is shown in Figure-\ref{645133}.