Deformable Neural Radiance Fields

model non-rigidly deforming scene

• warp each point to canonical 5d nerf
  • latent deformation code per image
  • estimate deformation params. v, s, t
• zeroing out hf for stable(non-overfitting) training
• elastic reg.: local jacobian should be close to a rotation matrix
• robust loss: refer to paper, read later
• bg: pre-assign some point to be stable
• positional encoding: low freq then high freq, anti-overfitting
  • NTK view of positional encoding: read later
• softplus for density, why?