目录

目录

GRAF: Generative Radiance Fields for 3D-Aware Image Synthesis


<GRAF> GRAF: Generative radiance fields for 3D-aware image synthesis
目录
  • 注:笔记在纸质版。待迁移电子版

Motivation

  • https://longtimenohack.com/posts/paper_reading/2020nips_schwarz_graf/58379603.png
  • While 2D generative adversarial networks have enabled high-resolution image synthesis, they largely lack an understanding of the 3D world and the image formation process.

    Thus, they do not provide precise control over camera viewpoint or object pose.

    因为2D GAN缺少对3D世界的理解;缺少图像生成过程的理解,所以不能提供对于camera viewpoint和物体pose的精确控制

  • 使用连续表征neural radiance filed
    • 从location x, view direction d映射到color c 和 体素密度 \(\sigma\)
  • 数据集使用unposed RGB images