510 likes | 636 Views
Using GenProg to Improve Computer Graphics. Adam Brady University of Virginia. 1988 . 1988 2009. Realistic materials are hard. Project Goal: How can we better model real materials in computer graphics?
E N D
Using GenProg to Improve Computer Graphics Adam Brady University of Virginia
Realistic materials are hard Project Goal: How can we better model real materials in computer graphics? Current modeling techniques fall under two distinct strategies
Strategy 1: Math Step 1) Think hard about physics
Strategy 1: Math Step 1) Think hard about physics Step 2) Channel James Clerk Maxwell
Strategy 1: Math Step 1) Think hard about physics Step 2) Channel James Clerk Maxwell Step 3) Write down an equation that approximates the physical world
Important Considerations What sort of patterns exist in the microscopic geometry structure? Can we statistically model light scattering? VERY FEW of these functions exist
Problems with This Strategy Finding fundamental approximations of the natural world is hard Assumptions limit accuracy:
Problems with This Strategy Finding fundamental approximations of the natural world is hard Assumptions limit accuracy:
Strategy 2: Measurement Step 1) Acquire a sample of a real material
Strategy 2: Measurement Step 1) Acquire a sample of a real material Step 2) Capture reflectance byphysically moving a light and camera
Strategy 2: Measurement Step 1) Acquire a sample of a real material Step 2) Capture reflectance byphysically moving a light and camera Step 3) Record data in one massive table
Problems with This Strategy • The table is HUGE • Millions of samples needed • No fundamental truths gained • Data only describes one specific material
We Need a Better Way • Math expressions are compact and contain adjustable parameters… • But inaccurate • Measured data is accurate.. • But HUGE and static
Project Outline • Use GenProgto modify reflectance functions • (Think small math expressions written in C)
Project Outline • Use GenProg to modify reflectance functions • (Think small math expressions written in C) • Measure fitness by rendering an image using the new function • Compare to reference image produced using measured data
Project Outline • Use GenProg to modify reflectance functions • (Think small math expressions written in C) • Measure fitness by rendering an image using the new function • Compare to reference image produced using measured data • Goal: Achieve measured data accuracy using a compact, adjustable math expression
INPUT EVALUATE FITNESS DISCARD ACCEPT OUTPUT MUTATE
INPUT Error: 0.4523
INPUT Error: 0.4523 MUTATE
INPUT EVALUATE FITNESS Compare Error: 0.4523 Error: ? MUTATE
INPUT EVALUATE FITNESS Compare ACCEPT Error: 0.4523 Error: ? Error: 0.3201 MUTATE
INPUT EVALUATE FITNESS Compare Error: 0.9702 DISCARD ACCEPT Error: 0.4523 Error: ? Error: 0.3201 MUTATE
INPUT EVALUATE FITNESS Compare Error: 0.9702 DISCARD ACCEPT Error: 0.4523 Error: ? Error: 0.3201 Error: 0.012 OUTPUT MUTATE
Key Differences from GenProg • Optimize for visual similarity rather than pass/fail correctness • Our fitness evaluation is continuous rather than discrete
Key Differences from GenProg • Optimize for visual similarity rather than pass/fail correctness • Our fitness evaluation is continuous rather than discrete • No fault localization • Entire C program can be changed
Key Differences from GenProg • We typically construct entirely new programs rather than make minor edits
Key Differences from GenProg • We typically construct entirely new programs rather than make minor edits • We use additional domain-specific mutation operators that operateat the expression level • sin(x+y) cos(x+y), vectora vectorb, etc.
Experiment Setup Learn a new function for a single material 32 generations, 2048 functions/generation
Results “I wouldn’t have thought of that!” -UVa Graphics Researcher
Current Work Results only consider a singlematerial Learning functions that fit multiple materials is much more difficult
Conclusions We used GenProglearn new compact, elegant expressions to accurately describe real materials However, additional work needs to be done to learn more general models