Deep Set-to-Set Matching and Learning