Paper Reading: Differential, Temporal Denoising and Control Variants


Differentiable Monte Carlo Ray Tracing through Edge Sampling

Work: compute rendering gradient

Application: Inverse Rendering(image->scene), deep learning

  1. model computation with step functions on different objects
  2. find and sampling on Edge to handle visibility issues
  3. integrate with PyTorch
  4. $$\nabla(s\cdot f)=(\nabla s)\cdot f + s \cdot (\nabla f)= \delta \cdot f + s \cdot (\nabla f)$$

Reparameterizing Discontinuous Integrands for Differentiable Rendering

Work: parameterize discontinuous, reduced noise of edge sampling

  1. use rotation to approximate displacement given by ray tracing: make the samples follow the geometry
  2. sample ray in convolution ways
  3. use control variates to reduce variance

Temporal Denoising

Sparsely Precomputing The Light Transport Matrix for Real-Time Rendering

Work: a real-time sampling framework that accelerates PRT precomputation

  1. separate dynamic lighting and static geometry inspired from PRT
  2. Why is sparse sampling fast?

Spatiotemporal Variance-Guided Filtering: Real-Time Reconstruction for Path-Traced Global Illumination

Work: a real-time reconstruction framework

  1. Temporal accumulation of frames in reconstruction
  2. Temporal antialiasing: create a color history buffer
  3. Temporal Variance estimation
  4. wavelet transform and edge-stopping

Control Variates

Find a integrand to redefine integral function.

$$\int f(x) dx = \int (f-g)(x) dx + G$$

Neural Control Variates

Work: use NN to generate a integrand for Control Variates

Image-space Control Variates for Rendering

Work: use control variate to give an optimal weight

Application: Gradient-domain path tracing, scene editing

  1. Control Variate to solve integral
  2. Optimal estimator weight
  3. use the weight in application

Leave a Reply

Your email address will not be published. Required fields are marked *