Learning Semantic Correspondence with Sparse Annotations