Towards Leveraging Sparse Infrared Datasets For Multiple View Synthesis, Few Shot Learning And Background Invariant Recognition