Exploiting appearance transfer and multi-scale context for efficient person image generation

Chengkang Shen, Peiyan Wang, Wei Tang

Abstract: Highlights•We introduce a novel two-stream context-aware appearance transfer network for efficient person image generation. It progressively transfers the appearance from the source stream to the target stream guided by their dense spatial correspondence and multi-scale context.•The proposed appearance transfer module is the first of its kind to use the target stream to query and transfer the source stream. It effectively handles the difficulty of large motion.•The proposed multi-scale context module is the first attempt to apply atrous convolutions for contextual modeling in person image generation. Multi-scale context helps the network recover occluded pixels.•Compared with state-of-the-art methods, our network achieves comparable or superior performance using much fewer parameters while being significantly faster. We also show our network has a great advantage when large pose transform occurs.