Interactive Pose Attention Network for Human Pose Transfer

In this paper, we propose an end-to-end interactive pose attention network (IPAN) to generate the person image in a target pose, where the generator of the network comprises a sequence of interactive pose attention (IPA) blocks to transfer the attended regions regarding to intermedia poses progressively, and retain the texture details of the unattended regions for subsequent pose transfer. More specifically, we design an attention mechanism by interacting with image and pose pathways to transfer the regions of interest based on the human pose, and capture the uninterested regions in the current IPA block against the uncertainty of the intermedia poses. In particular, we devise long-distance residual to inject the low-level features of the person image into the IPA blocks to keep its appearance characteristics. In terms of adversarial training, the generator exploits reconstruction loss, perceptual loss and contextual loss, and the discriminator exploits the adversarial loss. Quantitative and qualitative experiments conducted on the DeepFashion and Market-1501 datasets demonstrate the superior performance of the proposed method (e.g., FID value is reduced from 36.708 to 22.568 and 15.757 to 12.835 on Market-1501 and DeepFashion datasets, respectively).