-
Notifications
You must be signed in to change notification settings - Fork 1
Description
INFO
author
William Peebles1 Jun-Yan Zhu2 Richard Zhang3 Antonio Torralba4 Alexei A. Efros1 Eli Shechtman3
affiliation
1UC Berkeley 2Carnegie Mellon University 3Adobe Research 4MIT CSAIL
conference or year
CVPR 2022 - Oral Presentation
link
概要
- dense visual alignmentタスクにてGANで生成されたデータのみで学習.
- Inspired by the classic Congealing method, our GANgealing algorithm trains a Spatial Transformer to map random samples from a GAN trained on unaligned data to a common, jointly-learned target mode.
- 8つのデータセットにおいて既存のself-spervised learningを大きく上回る, supervised learningに匹敵ないし上回る性能を達成.
pair-wise alignmentではなくglobal joint alignment(データセット全体で画像のalignmentを行う)問題.joint alignmentされたデータセットで学習するとアライメントされていないデータで学習するよりも高品質の生成モデルを生成できるという報告もある(FFHQ, AFHQ, CelebA等)
提案手法
Tが変換するのが容易になるようなfixed latent vector
このままでは同じようなターゲットイメージを得るのには適していない.$G(c)$が$G(w)$の外観を維持しながら,姿勢と方向が同じターゲットを構築できることが理想なので,$w$を$c$の一部に一致するようにして学習.
実際,$mix(c, w)$はstyle mixingを行っている.$c$でラフにpose等をコントロール.$w$を後半のレイヤーに与えて,texture等をコントロール.
$ \bar{w} $はmean
LSUN Horsesは非常に多様な姿勢を持つデータセットであるため,クラスタリングを施して,それぞれのクラスタごとに$T$を学習させている.実画像でテストを行う際には,対応するクラスタを見つける必要がある.そこで,GAN Inversionなどの方法が考えられるが,ここでは単純な分類を行って該当するクラスタを判定している.分類機には$T$の重みを用いる (warpのためのheadをランダムに初期化したclassificationのheadに置き換えている).

検証
ablation study
新規性
In this paper, we showed that GANs can be used to train highly competitive dense correspondence algorithms from scratch with our proposed GAN-Supervised Learning framework.
議論,展望
Comment
date
30th May., 2022






