DUDE Deep Unsigned Distance Embeddings for Hi-Fidelity Representation of Complex 3D Surfaces

reconstruct 3D model using a unsigned distance function and a vector field

• both supervised
• since udf is not differentiable, approx. the “true surface” using normal estimation as a guide: local planary assumption
• min normal l2 loss up to module 180

  • loss being min of

    a - b

    and

    a - (-b)

• udf to meshees
  • subdivision, then marching cubes