Learning Multi-frame Visual Representation for Joint Detection and Tracking of Small Objects.