TOAC: Try-On Aligning Conformer for Image-Based Virtual Try-On Alignment

Recently, Image-based Virtual Try-on has garnered increasing attention within the realm of online apparel e-commerce, which aims to virtually superimpose garments onto images of portraits. Image-based virtual try-on generally consists of two steps: Image-based Virtual Try-on Alignment and Image-based Virtual Try-on Generation. In this paper, we focus on Image-based Virtual Try-on Alignment (IVTA), which plays a pivotal role in virtual try-on and aligns the target garment with the portrait. Current approaches for IVTA mostly adopt Convolutional Neural Networks (CNN) to extract local detailed features of both garments and portraits, ignoring the significance of global extensive features. To address this problem, we propose a novel model named Try-On Aligning Conformer (TOAC) to effectively aligns the target garment with the portrait and improve virtual try-on. Firstly, we integrate both Swin Transformer and CNN to comprehensively extract both global patterns and local details. Secondly, we propose a robust learned perceptual loss between generative reconstructed garment images and the ground truth to alleviate the overlap problem. Extensive experiments demonstrate the superiority of our proposed model compared to the state-of-the-art methods for virtual try-on alignment.