Convolutional Occupancy Networks

previous implicit models are prevented from structured reasoning

• only act in global scale -> overly smoothed reconstructions
  • positional encoding may help
• fc nets donot allow for integrating local features or translational equivariance
  • but, fc nets can achieve these from the smooth prior in MLP
• point -> feature -> proj to predefined plane -> refine feature
• extension: predefined plane -> dynamic plane
  • no significant performance gain compared to predefined plane
  • not to say with positional encoding introduced