Few-shot Semantic Encoding and Decoding for Video Surveillance