How can objects help action recognition?