← Context Encoders: Feature Learning by Inpainting