Precise Correspondence Enhanced GAN for Person Image Generation

To generate a realistic person image for pose-guided person image generation, especially for local body parts, is challenging. Two reasons account for it: (1) the difficulty for long-range relation modeling, (2) a deficiency in precise local correspondence capturing. We propose a Precise Correspondence Enhanced Generative Adversarial Network (PCE-GAN) to address these problems. PCE-GAN includes a global branch and a local branch. The former maintains the global consistency of the generated person image and the latter captures the precise local correspondence. More specifically, the long-range relation is well established via the spatial-channel Multi-layer Perceptrons module in the transformation blocks within both branches. The precise local correspondence is captured effectively by the local branch’s local-pair building and local-guiding modules. Finally, the outputs of each branch are combined for mutually improved benefits based on the enhanced correspondences. Experimental results show that, compared to previous state-of-the-art methods using the Market-1501 dataset, PCE-GAN performs quantitatively better, with a \(5.53\%\) and \(7.74\%\) improvement in SSIM and IS scores, respectively. Qualitative results for both Market-1501 and DeepFashion datasets are also provided herein to further validate the effectiveness of our method.