图像分割"LIP: Self-supervised Structure-sensitive Learning and A New Benchmark for Human Parsing"