Bringing Image Structure to Video via Frame-Clip Consistency of Object Tokens

On the other hand, one does often have access to a small set of annotated images, either within or outside the domain of interest. Here we ask how such images can be leveraged for downstream video understanding tasks.