DIST Rendering Deep Implicit Signed Distance Function with Differentiable Sphere Tracing

optimize forward and backward pass of neural rendering

• tldr, coarse-to-fine, aggressive ray marching
• apply to any sdf model
• sample per 4 pixel and double each 3 iterations
  • at 6 iter, back to origin res
• aggressive ray marching
  • x = x + 1.5 * sdf(x) * rd
  • sample points under surface
• converge if a pixel donot confuse with its neighbor
  • refer to supp. for details
• gradient bp
  • tbh, i donot understand it