DISN Deep Implicit Surface Network for High-quality Single-view 3D Reconstruction

limitations

• low res.
• fixed representation but limied topology(e.g. holes)
• Chanfer Distance or EMD: approx. metrics
• study global features but ignore details
  • introduce a local feature module
  • project query point into input image -> get a local patch

Signed Distance Functions

• implicit representation
• propose a DISN to predict SDF
• potential to generate infnite res. 3D model
• continuous function: R^3 -> R, SDF(p) = distance from surface

DISN

• regress SDF
  • based on global feature and local feature
• local feature?
  • extracted from endoder, then concat
• map 3d point to 2d image
  • estimate cam pose first!
• marching cubes to get iso-surface