Project 4: Utkarsh Dwivedi

[utkarshdwivedi3997]

Repo Link
Features

• G-buffer for position, normal, depth
• ATrous edge-avoiding wavelet transform denoiser

Feedback

• The amount of time given to implement a lot in project 3 and the amount of time given to implement this incredibly simple, and outdated, denoiser, is extremely unbalanced.
• The code base and instructions need to be updated and is simply wrong in a few places:
  • the gbufferToPBO() kernel incorrectly multiplies the normalised 0 to 1 t value by 256.0f. This is beyond the colour range and should be 255.0f.
  • the normals buffer that is shown in the instructions uses abs(normal) instead of the correct way of visualising normals (normal + 1) / 2. Taking an abs makes walls pointing in either left or right directions appear red. This is incorrect and threw me off for a bit while I was working on this assignment, especially since this is coming directly from the image in the instructions!
  • The source code in the paper has the following line when calculating normal weightage: dist2 = max(dot(t,t)/(stepwidth*stepwidth),0.0); but I found that changing it to simply dist2 = dot(t,t); yields MUCH better results.
• This is a very old paper that describes a very outdated denoising technique, and doesn't even have the need to use any of CUDA's advanced optimisation features. This assignment should be completely replaced with something more up-to-date.
• The results of this technique are not impressive for the current times. The recommendation to integrate it into our path tracers should be removed, because with complex materials it simply made my renders look worse.