Typical input for 3D reconstruction: images and point clouds

Task: converting rough, imprecise, (distorted compared to gt) sketch to clean, complete 3D shape

  • rough -> requiring interpretation of noisy signals

High quality 3D modeling:

  • high resolution
  • sharp features

1. 2D sketch

  • imcomplete -> requiring hallucination of missing parts


  • ambiguities ->
    • hand-desinged priors
    • limiting applications
    • learn the shapes from data
    • implicitly inferring relevant priors

2. 3D sketch

3. Loss Functions

  • 2D: use differentiable renderer and measure inconsitencies in 2D
  • 3D: Chamfer / regularized Wasserterin

4. Terminology


results matching ""

    No results matching ""