58 lines
No EOL
1.4 KiB
Markdown
58 lines
No EOL
1.4 KiB
Markdown
# Notes diffable Monte carlo RT
|
|
|
|
## Raytracing formula
|
|
|
|
- geometry term discussed later
|
|
- Emission + All light reflected towards point
|
|
- Yields infinite recursion -> not calculable
|
|
|
|
## Visualization
|
|
|
|
- Explain image
|
|
- No indirect lighting!
|
|
- Output image is what we would expect (explain shade)
|
|
|
|
## Differentiable rendering
|
|
|
|
- That function is dependent on renderer
|
|
- Renderer needs to be differentiable
|
|
|
|
## Importance
|
|
|
|
- Inversely render complex indoor scenes
|
|
- "Fool" neural network
|
|
- Real time realistic shading in AR
|
|
- Application in maritime research
|
|
|
|
## Adversarial image generation
|
|
|
|
- Example for classification on slide 2!
|
|
- Fool neural netweork into wrongly classifying input data
|
|
- Optimize Image into wrong class
|
|
|
|
## Why differentiable rendering is hard
|
|
|
|
- Example later
|
|
- geometry term explanation later
|
|
|
|
## Former methods visualization
|
|
|
|
- Plane lit by a point light source.
|
|
- gradient with respect to the plane moving right
|
|
- light source remains static => the gradient should only be $\ne 0$ at the boundaries
|
|
- OpenDR and Neural not able to correctly calculate the gradients
|
|
- they are based on color buffer differences
|
|
|
|
## Edge sampling
|
|
|
|
- Approximate point lights using small area lights
|
|
- Specular => angle of incidence = angle of light reflected
|
|
- only lambertian materials
|
|
|
|
## Edge Sampling - Math Background
|
|
|
|
- Heaviside step functions in $f_i(x,y)$
|
|
|
|
## Inverse Rendering - Results in this paper
|
|
|
|
- ADAM: talk by Mr. Wu |