Paper

• Published in 2019
• ICCV paper

Introduction

• GAN Network를 이용한 Disentangled 3D Representation learning
• Unsupervised learning(Pose labelled data X)

Generator Network

Conv → DeConv

1. Learning 3D Representation

   • Generates 3D representations from a constant
   • AdaIn(Adaptive instance normalization) 사용 - z adapt 한다.

2. Learning with view-dependent mappings(3D Transform)

   • Random pose transformation - Crucial in disentanglement(타 논문에서 확인)

   • Rigid body transform(Rotation only - elevation, azimuth), trilinear resampling(scaling)

     범위 : elevation(70 ~ 110) , azumith(220~320)

3. Projection Unit

• Two layers for the network to learn projection

• Projection by Reshaping (3D features → 2D features)

  ex) W×H×D×C → W×H×(D·C)

4. Fake Image Generation

• Generate fake image from extracted features

Discriminator Info

• Uses instance normalization & spectral normalization

  • Spectral Normalization(Discriminator 성능 향상을 위해 Lipschitz 상수를 제한)

    Weight matrix를 Matrix의 가장 큰 e.v로 나눈다