Papers of the day   All papers

The Surprising Effectiveness of LinearUnsupervised Image-to-Image Translation

Comments

Phillip Isola: Surprising and fun result: Unpaired image translation without a deep net, just a _linear_ transformation: https://arxiv.org/abs/2007.12568 (and no GAN too!)

4 replies, 329 likes


andrea panizza: Another surprising paper by Eitan Richardson & Yair Weiss https://arxiv.org/abs/2007.12568. TL;DR: simple linear method beats SOTA methods for img2img translation such as CycleGAN, succeeding where they fail, and doing just as well where they succeed. It raises some questions....

5 replies, 289 likes


AK: The Surprising Effectiveness of Linear Unsupervised Image-to-Image Translation pdf: https://arxiv.org/pdf/2007.12568.pdf abs: https://arxiv.org/abs/2007.12568 https://t.co/TbuyDS2lEM

0 replies, 62 likes


will grathwohl: powerful stuff...need more of this. great follow up to their prior "you can do just as good as GANs a MOG" paper (http://papers.nips.cc/paper/7826-on-gans-and-gmms) adds more kindling to the "everything I do is bullshit" fire. plz burn me down and free me from this life <3

1 replies, 15 likes


Daisuke Okanohara: Unsupervised image-to-image translation (e.g., CycleGAN) often uses DNN. However, (surprisingly) linear orthogonal translation can represent many of them, and such translation can be obtained by just PCA and correspondence findings. https://arxiv.org/abs/2007.12568

0 replies, 15 likes


arXiv CS-CV: The Surprising Effectiveness of Linear Unsupervised Image-to-Image Translation http://arxiv.org/abs/2007.12568

0 replies, 6 likes


Artsiom Sanakoyeu: Interesting

0 replies, 3 likes


Kevin Yang 楊凱筌: 1. Apparently it's possible to learn unsupervised image-to-image translation (eg colorization)?! 2. And a linear encoder-decoder model is really good at it! Eitan Richardson, Yair Weiss https://arxiv.org/abs/2007.12568 https://t.co/VB8SGZFvvq

0 replies, 1 likes


Andrew Carr: L I N E A R > ~ L I N E A R https://arxiv.org/abs/2007.12568 1000x faster? Yes please. Unsupervised image to image translation, locality bias free. Next step, videos https://t.co/rJdkvauvbf

0 replies, 1 likes


akira: https://arxiv.org/abs/2007.12568 Domain transformation by GAN relies on localized bias, and even simple entire image change tasks (e.g., flipping up and down) fail. They found that even linear transformations alone can work well for those tasks. https://t.co/AujmcnkB1y

0 replies, 1 likes


Content

Found on Jul 27 2020 at https://arxiv.org/pdf/2007.12568.pdf

PDF content of a computer science paper: The Surprising Effectiveness of LinearUnsupervised Image-to-Image Translation