Haar-wavelet based texture inpainting for human pose transfer

Abstract

Human Image Generation is an important information processing technique. Producing realistic-looking texture is crucial for Generative Adversarial Networks (GAN) based person image generation. Existing methods follow a “downscale-upscale” strategy that source images are usually downscaled to extract features and saved the cost of storage. Meanwhile, the upscaling process is applied to recover the details based on these features. The loss of high-frequency components during the downscaling process, however, is in accordance with the Nyquist-Shannon sampling theorem, which creates an ill-posed difficulty during the upscaling process. In this paper, we design a Haar-wavelet based texture inpainting network (HWTIN) to mitigate the ill-posed problem in pose transfer task. In the downscaling process, to divide the source image into high-frequency and low-frequency contents, we construct a Haar-based Wavelet Module (HBM). In this way, We can preserve these high-frequency information in the generation process. We also design an inverse HBM (IHBM) to utilize these high-frequency information in the upscaling process. Extensive results on mainstream datasets demonstrate that HWTIN outperforms state-of-the-art (SOTA) methods quantitatively.