Selective Restructuring of Bo nding Vol me Hierarchies for Bounding Volume Hierarchies for Dynamic Models y Sung-Eui Yoon KAIST (Korea Advanced Institute of Science and Technology) and Technology)
At Previous Class At Previous Class ● Studied multi-resolutions, culling, cache- Studied multi resolutions culling cache coherent layout techniques What is one of major problems of these techniques? 2
Motivations Motivations ● Dynamic scenes are widely used Dynamic scenes are widely used ● Movies, VR applications, and games ● Complex and large dynamic scenes ● E g high resolution explosion tears and ● E.g, high-resolution explosion, tears, and fractures 3
An Example of Exploding Dragon (252K triangles) (252K triangles) 4
Ray Tracing Dynamic Scenes Ray Tracing Dynamic Scenes ● Acceleration hierarchy construction A l ti hi h t ti ● e.g., kd-trees, bounding volume hierarchies, grids etc grids, etc ● Hierarchy traversal ● Hierarchy traversal ● Perform ray-triangle intersection tests ● Key issue ● Update the hierarchy as triangles deform ● Update the hierarchy as triangles deform 5
Bounding Volume Hierarchies (BVH) based Ray Tracing (BVH) based Ray Tracing ● Employed early in [Whitted 80] E l d l i [Whitt d 80] ● kd-trees and grids became popular for static models in 90’s models in 90 s ● Recently get renewed interest in ray ● Recently get renewed interest in ray tracing dynamic scenes [Wald et al. 07, Lauterbach et al. 07, Larsson et al. 03] Lauterbach et al. 07, Larsson et al. 03] ● Simple, but efficient BVH update method is available ● Can have better performance 6
BVHs BVHs ● Object partitioning hierarchies Object partitioning hierarchies ● Uses axis-aligned bounding boxes ● Considers surface area heuristic (SAH) ● Considers surface-area heuristic (SAH) [Goldsmith and Salmon 87] A BVH A BVH A BVH A BVH 7
Two BVH Update Methods Two BVH Update Methods BV refitting BV refitting • O(n) O(n) O(n) O(n) Frame 1 Frame 1 • Poor Poor- -quality BVs quality BVs Frame 2 Frame 2 BV reconstruction BV reconstruction • O(n log n) O(n log n) • Good Good- -quality BVs quality BVs 8
Our Goal Our Goal ● Existing BVH update methods Existing BVH update methods ● Work at particular classes of dynamic scenes ● Design a robust BVH update method Design a robust BVH update method ● Works well with wide classes of dynamic scenes ● I mproves the performance of ray tracing ● I mproves the performance of ray tracing 9
Our Contributions Our Contributions ● Proposes a novel algorithm to selectively Proposes a novel algorithm to selectively restructure BVHs [Yoon et al., EGSR 07] ● Selective restructuring operations ● Selective restructuring operations ● Two probabilistic metrics: culling efficiency and restructuring benefit g Refit Refit Restructure Restructure BVH 10
Example of Exploding Dragon Model Dragon Model Ray tracing time (sec): Ray tracing time (sec): # of intersections # f i t ti construction + traversal BV refitting Complete Selective reconstruction restructuring restructuring 11
Runtime Captured Video – BART Model (65K triangles) BART Model (65K triangles) ● Compared with the BV refitting method Compared with the BV refitting method E Enabled primary Enabled primary E bl d bl d i i & shadow rays & shadow rays Single thread Single thread 12
Probabilistic BVH Metrics for Ray Tracing for Ray Tracing ● Culling efficiency Culling efficiency ● Quantifies the quality of any sub-BVHs ● Measures the expected # of intersection tests ● Measures the expected # of intersection tests for a ray ● Restructuring benefit ● Predicts the performance improvement Predicts the performance improvement ● Measures improved culling efficiency when restructuring sub-BVHs 13
Culling Efficiency Metric Culling Efficiency Metric ● Measure the expected # of intersection Measure the expected # of intersection tests for a ray ● Measured in a view-independent manner ● Measured in a view-independent manner ● Recursively computed with child nodes considering SAH [Goldsmith and Salmon 87] g [ ] Ray Ray BVH 14
Validation of Culling Efficiency Metric Metric High correlation! A good metric measuring the quality of BVHs 15
Restructuring Benefit Metric Restructuring Benefit Metric ● Predicts improved culling efficiency when Predicts improved culling efficiency when restructuring sub-BVHs ● Should not perform actual restructuring ● Should not perform actual restructuring ● Restructure the sub-BVHs ● Restructure the sub BVHs ● Only if the restructuring benefit is bigger than the restructuring cost g 16
Major Observation Major Observation ● Restructuring two nodes with BV overlaps Restructuring two nodes with BV overlaps can improve the culling efficiency ● Assumes that restructuring operation will ● Assumes that restructuring operation will remove all the BV overlaps A BVH A BVH 17
Selective Restructuring Operations Restructuring Operations 18
Validation of Restructuring Benefit Metric Restructuring Benefit Metric ● Compare the expected values against the Compare the expected values against the observed values ● 80% of the observed values are 25% off from ● 80% of the observed values are 25% off from the expected values 19
Overall Framework Overall Framework ● At a new frame At a new frame ● Refits BVs with deformed triangles ● Performs our selective restructuring algorithm ● Performs our selective restructuring algorithm ● Runs BVH-based ray tracing 20
Detecting BV Overlaps Detecting BV Overlaps ● Brute-force method Brute force method ● Requires O(m 2 ) where m is # of BVs ● Hierarchical traversal and culling Hierarchical traversal and culling ● I nspired by efficient collision detection methods methods 21
Overview of Selective Restructuring Algorithm Selective Restructuring Algorithm ● Hierarchical refinement phase Hierarchical refinement phase ● Restructuring phase 22
Overview of Selective Restructuring Algorithm Selective Restructuring Algorithm ● Hierarchical refinement phase Hierarchical refinement phase ● Detects nodes with BV overlaps during hierarchy traversal hierarchy traversal ● Restructuring phase 23
Overview of Selective Restructuring Algorithm Selective Restructuring Algorithm ● Hierarchical refinement phase Hierarchical refinement phase ● Restructuring phase ● Restructure node pairs with higher benefits in a R t t d i ith hi h b fit i greedy manner 24
Evaluating Our Algorithm Evaluating Our Algorithm ● I mplement BVH-based ray tracer I mplement BVH based ray tracer [Lauterbach et al. 06] ● Tests with four dynamic scenes having different ● Tests with four dynamic scenes having different characteristics 25
● Cloth simulation (92K) Cloth simulation (92K) Dynamic Scenes Dynamic Scenes 26
● N-body simulation (146K) N body simulation (146K) Dynamic Scenes Dynamic Scenes 27
● BART ● BART (65K) Dynamic Scenes Dynamic Scenes ● Exploding dragon ● Exploding dragon (252K) 28
Prior Works Prior Works ● BV Refitting [Wald et al. 07, Bergen 97] BV Refitting [Wald et al 07 Bergen 97] ● Complete re-construction from scratch ● Other two hybrid methods ● Based on a simple heuristic ● RT-Deform [Lauterbach et al. 06] RT D f [L t b h t l 06] ● LM method [Larsson and Akenine-Möller 06] 06] 29
Performance Improvement Ratio Performance Improvement Ratio Complete Refitting re-construction only Exploding 8.5 11 dragon g N-body 1.8 > 80 simulation simulation BART 1.1 28 Cloth 4.7 0.96 simulation simulation 30
Image Shots from Cloth Simulation Simulation Initial frame 31
Performance Improvement Ratio Performance Improvement Ratio Robust performance improvement Robust performance improvement Robust performance improvement Robust performance improvement across our benchmarks across our benchmarks Complete Complete Refitting Refitting RT- RT LM LM Deform method const. only Exploding Exploding 1.65 1 65 2 16 2.16 8.5 11 dragon N body N-body 1.25 1 25 1.36 1 36 1.8 1 8 > 80 80 simulation 2.5 2.5 1.11 1.11 BART BART 1 1 1.1 28 28 Cloth 1.03 1.29 4.7 4.7 0.96 0.96 simulation i l ti 32
Conclusions Conclusions ● Novel algorithm to selectively restructure Novel algorithm to selectively restructure BVHs ● Based on selective restructuring operations and ● Based on selective restructuring operations and two BVH metrics ● Has more robustness and deals with bigger gg scene complexity ● Can be used in other applications - Dynamic scenes are available Dynamic scenes are available 33
Will study collision detection ● Will study collision detection At Next Class At Next Class 34
Recommend
More recommend
