gangealing

Why we included this project

Anyone who works on dense correspondence will find this a solid reference implementation of a CVPR 2022 oral paper. GANgealing trains a Spatial Transformer on synthetic GAN images to warp unaligned inputs into a shared canonical view, and that alignment carries over to real photos and video without extra supervision. The per-pixel matches it produces support object tracking and label propagation, and the included Colab notebook shows how to propagate drawn edits across frames realistically. With pre-trained models for eight datasets and the training and evaluation code plus visualization utilities, you can run it out of the box or use it as a base for reproducing and extending the method.