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