Parallax-Interpolated Shadow Map Occlusion Pismo Parallax-Interpolated Shadow Map Occlusion Ivan Neulander
Introduction: Soft Shadows Problem: Render soft shadows from an area light source Light source can be any polygonal model Want fast solution, of sufficient quality for visual effects Want speed vs. quality control Area Light Penumbra Umbra
Introduction: Soft Shadows Some Solutions: Ray tracing Very high quality, physically accurate but usually slow “Shadow Map” ray tracing: project ray to map Lower quality, relatively fast Samples shadow map many times per ray Soft Shadows by Ray Tracing Multilayer Transparent Shadow Maps: Xie, Tabellion, Pearce Efficient Image-Based Methods for Rendering Soft Shadows: Agarwala et al Use many shadow maps: overlapping shadows Fast at low quality but very slow at high Use many shadow maps: Pismo Almost as fast as (c), but much higher quality Samples shadow map few times per ray 128 Shadow Maps Pismo with 4 Shadow Maps 64 Shadow Maps 256 Shadow Maps 16 Shadow Maps 4 Shadow Maps 8 Shadow Maps 32 Shadow Maps
Soft Shadow Example Area Light
Hard Shadows from 3 Point Lights with Shadow Maps
Pismo Shadow Estimation: Problem Problem: Estimate shadow ray PL, extending from P to arbitrary point L on area light Use the 3 shadow maps rendered from the triangle vertices, all fixated on a common focal point F P L
Pismo Shadow Estimation: Solution Solution: 1) Compute position of putative occluder O Displace P toward L by distance d How to compute d ? P d O L
Pismo Shadow Estimation: Solution Shadow Map #1 Solution: 2) Project O into screen space of Shadow Map #1 Use depth value at this pixel to compute ray occlusion Using O instead of P accounts for parallax shift between view from Shadow Map camera and view from L P d O L
Pismo Shadow Estimation: Solution Shadow Map #2 Likewise, project O into screen space of Shadow Map #2 P d O L
Pismo Shadow Estimation: Solution Shadow Map #3 Likewise, project O into screen space of Shadow Map #3 P d O L
Pismo Shadow Estimation: Solution Blending the Three Shadow Maps Barycentrically blend the 3 shadow maps’ occlusion results to get the final occlusion estimate for the shadow ray PL Uses 3 shadow map queries per ray P d O L
Pismo Shadow Estimation: Solution Compute Multiple Shadow Rays Repeat the process for multiple shadow rays between P and other light positions on the area light P
Estimating Initial Depth d: Preprocess Shadow Maps For each Shadow Map, we generate a Bled Shadow Map: an image-processed copy where we bleed out silhouette pixels apply a slight blur One-time preprocess per shadow map Regular Shadow Map stores multiple occluder depths per pixel (layered) Bled Shadow Map stores max depth only
Estimating Initial Depth d: Compute d1 Project P to a pixel in Bled Shadow Map #1 If queried depth is less than depth from P to L, subtract and we’re done; else, compute following average: Sample multiple jittered pixels in (layered) Shadow Map, centered about original pixel For each sample: If queried depth is less than depth from P to L, add difference to average Computed once per shading sample (not per ray) P d1 L
Estimating Initial Depth d: Compute d2 Project P to a pixel in Bled Shadow Map #2 Estimate value for d2 as above Computed once per shading sample P d2 L
Estimating Initial Depth d: Compute d3 Project P to a pixel in Bled Shadow Map #3 Estimate value for d3 as above Computed once per shading sample P d3 L
Estimating Initial Depth d: Blend d1, d2, d3 Barycentrically blend d1,d2,d3 to get d for given L Use blended d value to compute O as shown Computed per-ray (but no shadow map queries) P d L
Depth Compensation: Problem Depth from L to P (LP) differs from depth from Shadow Map D to P (DP) Need to compensate depth estimates that use D in place of L P D DP LP L
Depth Compensation: Solution Adjust initial estimates d1,d2,d3 by LP – DP Blend d1,d2,d3 and compute O Adjust depth queried from D (based on O) by LO – DO D (Shadow Map) F (Focal Point) L LO LP O P DO DP
Contact Shadows Problem: Pismo requires a higher shadow bias than regular shadowing to avoid self-shadow artifacts This attenuates crisp short-range shadows Solution: Compute contact shadows at close range and blend them in with Pismo Shadows
Contact Shadows: Implementation A Contact Shadow: is based on hard shadows from discrete shadow maps consists of 2 or more such shadows that overlap fades out beyond Pismo shadow bias distance is blended with Pismo shadow over bias distance blend =
Results: Area Light with 4 Shadow Maps Pismo Off Pismo On 128 rays / pixel Area Light
Results: Area Light with 5 Shadow Maps Pismo Off Pismo On 128 rays / pixel Area Light
Performance: Pismo vs. Ray Tracer 1600x1200 resolution 128 rays / pixel Athlon64 @ 2800 MHz Performance: Pismo vs. Ray Tracer Pismo@4 5 DMs: 33 s Final: 263 s Total: 296 s Pismo@5 5 DMs: 42 s Final: 246 s Total: 288 s Ray-Traced Total: 1073 s
Quality: Large Area Light Ray Tracer 9 Blurred Shadows 4 Blurred Shadows 5 Blurred Shadows 9 Pismo Shadows 4 Pismo Shadows 5 Pismo Shadows 4 Hard Shadows 5 Hard Shadows 9 Hard Shadows
4 Hard Shadows 5 Hard Shadows 9 Hard Shadows 9 Blurred Shadows 4 Blurred Shadows 5 Blurred Shadows 9 Pismo Shadows 4 Pismo Shadows 5 Pismo Shadows
Quality: Pismo@4 vs. Ray Tracer Pixel Difference Ray Tracer
Quality: Pismo@5 vs. Ray Tracer Pixel Difference Ray Tracer
Quality: Pismo@9 vs. Ray Tracer Pixel Difference Ray Tracer
Conclusions Pismo is a useful enhancement to rendering soft shadows with multiple shadow maps Conceptually simple & easy to implement integrates easily into an existing rendering pipeline Provides an intuitive control over speed/quality Creation of multiple shadow maps easily accelerated: Multiprocessing GPU rendering Morphing
Future Work Accelerate creation of shadow maps Automate tessellation of light polygons Improve blending of contact shadows Parallelize Pismo sampling
Ivan Neulander www.rhythm.com/~ivan/pubs.html ivan@rhythm.com