Occupancy Networks Learning 3D Reconstruction in Function Space

3D output representation

• both memory efficient and efficiently inferred from data
• voxel: memory hungry
  • even using octrees: 256^3
• point clouds: need additional processing steps
• mesh: hard t oinferred by feed-forward network
• Occupancy Networks
  • field similar to texture field
  • loss: cross-entropy
  • MISE: incrementally building an octree
  • refine using gradient information: gradient toward GT normal
• experiment: similar to texture field