Guided Visibility Sampling

1
Guided Visibility Sampling

• Peter Wonka, Michael Wimmer, Kaichi Zhou, Stefan
  Maierhofer, Gerd Hesina, Alexander Reshetov

2
Problem Statement

• Input Triangulated Model, Region
• Output Triangles visible from input region

Model
Input Region
Visible
Invisible
3
Practical Application

• Preprocessing Stage
• Locate all possible viewing cells
• For each cell, compute potentially visible set
• Rendering Stage
• Identify current viewing cell
• Only render visible set of triangles

4
Impetus for Sampling

• Pros
• Quick
• Simple
• No overestimation
• Progressive
• Cons
• Incorrect when undersampling

5
Naïve Sampling
Object Plane
Camera Plane
6
Guided Visibility Sampling Can they beat random?

• Guided Visibility
• Sampling

Uniform Random Sampling
7
Guided Visibility Sampling Algorithm

• Step 1 Pick Random Rays

Object Plane
Camera Plane
8
Guided Visibility Sampling Algorithm

• Step 1 Pick Random Rays

Object Plane
Camera Plane
9
Guided Visibility Sampling Algorithm

• Step 1 Pick Random Rays

• Step 2 Place all hit triangles in a queue

Object Plane
Camera Plane
10
Guided Visibility Sampling Algorithm

• Adaptive Border Sampling
• Guide rays that hit new triangles
• perturb endpoint to locate neighbor triangles
• Repeat for all new triangles discovered this way

11
Guided Visibility Sampling Algorithm

• Reverse Sampling
• Explore triangles that are not adjacent
• perturb ray start point to fill gaps in
• Repeat for all new triangles discovered this way

Front View
Top View
12
Adaptive Border Sampling

• Subdivision

13
Adaptive Border Sampling

• Enlarging Triangles
• Grow Triangle numerical problems
• Sollution
• Ennaegon
• 9-sided
• Computed in Ray Space to benefit edge-on triangles

14
Adaptive Border Sampling

• Compute Ennaegon in ray space to deal with
  edge-on triangles
• Top View Top View

15
System Overview

• ray hits old triangle

16
Results
17
Conclusions

• Sampling vs Exact/Conservative
• Contributions
• Large models
• No restriction on input scene
• Few image errors
• Problems
• Ray tracing w/o locality
• Ray tracing precision
• No error bounds

18
Notes

• Large models are tiny
• In core algorithm
• Deals poorly with T-junctions
• Too many tweakables
• Max vs Ave error
• Could result in temporal aliasing?