Spmpg: Robust Person Image Generation With Semantic Parsing Map

Person image generation is an interesting research topic that attracts many attention. In this paper, a robust body semantic parsing map guided automatic people pose synthesizing approach is proposed. Compared with commonly used skeleton based generation, our approach takes full advantage of available human body information, and hence successfully achieves improved performance. In the proposed method, extensive attention and multiscale discrimination mechanisms are enrolled to further advance generation performance. Verification experiments are extensively conducted on the Market1501 and Deepfashion dataset to systematically evaluate the effectiveness of our approach.