Image synthesis and OpenGL: post-processing

1. Principle
2. Gamma correction
3. Color correction and clamping
4. Blending colors / Fog
5. Depth of field
6. Motion blur
7. Lens flare
8. Glowing
9. Tone mapping
10. Screen Space Ambient Occlusion
11. Antialiasing
12. Sources
    

Principle

1. Draw the 3D scene inside a texture using FBOs
2. Map the texture on a quad that covers the entire screen
   
3. Apply the post-process in a fragment shader 

• OpenGL / GLSL can basically be used to perform all kind of image processing!
• Usually much faster than applying the filter on the CPU
  
• Everything works in parallel!
  

Gamma correction

• The screen does not render colors linearly
• We need the "inverse function" to correct colors
  
• The function "looks like" a power function of 2.2
• But: more complex solutions exist to precisely reproduce colors (necessitate mesuring screen responses)
  

Color correction and clamping

• Modify Hue, saturation and brightness to obtain the desired effect
  

Blending colors / Fog

• Blend between the scene color and the background color according to geometric information (depth)
• The blending factor may be computed using multiple ways:
• linear: \( f = (end-z)/(end-start) \)
• exp: \( f = e^{-(density*z)} \)
• exp2: \( f = e^{-(density*z)^2} \)

• Blending can used between colors but also textures using any kind of function to obtain the desired result
  

Depth of field

• Blur the rendering
• using any kind of filter (mean, gaussian, etc)
• based on a geometrical property (depth)

Motion blur

• Blur in a specific direction
• Can be given by the camera motion
  

Crytek interview

Crysis 2

Lens flare

• Blend with light textures
  
• Need specific operations to check whether the light/sun is visible

Glowing

• From the rendering (a)
• Render the glowing parts in a separate buffer (b)
• Apply a blur on this buffer (for instance using separate gaussian convolution) (c)
  
• Add this buffer to the original one (d)
  

Tone mapping

• Goal: remap values of a HDR image (comprised in the range [0,inf]) to the range [0,1]
• Many methods: global and local
• Usually need some global information like min/mean/max luminance of the image
• How can we efficiently compute these values on the GPU? 
• Solution: manually create mipmaps of the given image
• At each image resolution
• Look at the four corresponding pixels in the highest resolution
• keep only the min/max/addition of values
• The last level contains the needed values
  

Original HDR Image (cannot be displayed properly)

Tone mapped (LDR) Image

Screen Space Ambient Occlusion

• Goal: approximate Ambient Occlusion in screen space
• Based on the depth map (but many methods that can also use normals/noise/etc)
• Basic idea: if a neighbor of a pixel is farther than the pixel itself, it has big chances to be in the shadow
• More information here / here / here / here / etc
  

SSAO	Comparison with/wihout SSAO

• Independent from geometry
• Dynamic
• No pre-processing
  

Antialiasing

• MLAA (Morphological Anti-Aliasing)
• blend colors around silhouettes infered from pixel discontinuities
• SMAA (Subpixel Morphological Anti-Aliasing)
• GPU friendly
• Again: many appearing methods 
• More information in this very good Siggraph Course
  

MLAA

SMAA

Sources

• J. Baril, F. Berthaut, N. Mellado