1. Introduction

• Nerf
  • Models a scene with radiance field and density with weights of NN
  • Use volume rendering to synthesize new views
  • shortcomings
    • Controlled setting에서만 잘 작동함
    • lighting, content is all static due to short time
    • performance degrades with moving object, variable illumination
  • Assumes static world
    • the density and radiance of the world is constant → any photos taken from same position, orientation must be identical
  • Real world is not static
    • same location → people move, sky, sun, weather etc
    • same location, same time → exposure, color correction, camera etc

2. NeRF in the Wild

• NeRF’s assumption
  • that a point in 3D space observed from the same position and viewing direction in two different images has the same intensity.
• Violation of Assumption
  • Photometric variation
    • illumination variance
      • time of day, weather conditions etc
      • Photographic imaging pipelines
  • Transient Objects
    • Never isolated
    • cars, objects, pedistrians etc
• NeRF-w resolves these issues by
  • Latent Appearance Modeling
  • Transient Object Disentanglement

2-1. Latent Appearance Modeling

• Each image is assigned an appearance embedding vector by adopting GLO

• Guaranteed 3D geometry is static and shared across all images
• Freedom to vary emitted radiance of the scene in particular image

2-2. Transient Object Disentanglement

• Transient Head