End-to-End Learning of Geometric Deformations of Feature Maps for Virtual Try-On

Pytorch implementation of WUTON: a Warping U-net for a Virtual Try-On system using an agnostic person representation along with it's densepose information.

Paper

Output Results

• The first one is the appaerl image cut out of the person image.
• The second one is the product image of the apparel after geometric transformation based on predicted TPS transformation parameters.
• The third one is the person image.
• The fourth one is the person image reconstructed from an agnostic image, given the apparel product image.
• The sixth one is of the same person wearing a differnt apparel altogether which is the fifth image.

Architecture of the Network

Prerequisites

• Linux
• Python3, PyTorch
• NVIDIA GPU (8G memory or larger) + CUDA cuDNN

Dataset

We have used Person Image along with it's apparel image scapred from zalando website. Along with Person-Apparel pair we have used the segmentation information and desnepose information of the person to train our model.