Single Sample Soft Shadows Using Depth Maps

1
Single Sample Soft Shadows Using Depth Maps

• Stefan Brabec and Hans-Peter Seidel
• Max-Planck-Institut für Informatik

2
Overview

• Introduction
• Single sample soft shadows withdepth maps
• Implementation
• Results
• Conclusion Future Work

3
Introduction

• Good-looking shadows greatly improve quality of
  rendered images
• For interactive applications mostly hard-shadows
  are used

4
Hard vs. Soft Shadows

• Hard Shadows
• Point light / directional light
• Fast rendering
• Shadow Volumes, Shadow Mapping
• Soft Shadows
• Area light / volume light
• Slow rendering
• Sampling the light source

5
Soft Shadows

• E.g. sampling the area light source
• Needs a number of hard-shadow calculations and
  accumulates the results
• More samples gives smoother penumbra

many samples expensive rendering
6
Soft Shadow Example

• Ground plane shadows

1. Initialize FB (white) 2. For each sample point
do 2a. Render scene 2b. Subtract 1/N from
FB only once for each pixel
(stencil) !
Image from ATI Developers Site
7
Parkers Method

• Soft shadows for point lights
• Adding an outerhull to an object
• Rays inside outerhull are attenuated

8
Parkers Method

• Each ray is attenuated according to distance to
  inner object

9
Parkers Method

• Pros
• Minimal overhead for ray casting(speed
  implementation work)
• Cons
• Not suitable for hardware rendering
• Adds only an outer penumbra very thick shadows

10
Our Method

• Parkers idea adapted for hardware rendering ?
• Use sampled input data (depth maps)
• Shadow map algorithm
• Concept of extended occluders
• Search for hard shadow/lit regions
• But
• Penumbra for outer and inner regions

11
Shadow Mapping

• Render depth map from light position
• Render from camera and comparetransformed z
  coordinate with depthstored in lights depth map

12
Shadow Mapping
13
Computing Hard Shadows

• Given two depth maps (light cam)we compute
  shadows as

foreach(x,y) P (x,y,depth_camerax,y) P
warp_to_light(P) if( depth_lightPx,Py lt Pz
) pixel is blocked else pixel is lit
14
Adding Outer Penumbra

• Adding an outer penumbra regionto this hard
  shadow

foreach(x,y) P (x,y,depth_camerax,y) P
warp_to_light(P) if( depth_lightPx,Py lt Pz
) pixel is blocked else pixel is lit
lit or outer penumbra ?
15
Adding Outer Penumbra

• Add penumbra by searching forthe nearest hard
  shadow
• Use (Px,Py) as the starting point
• Search neighborhood fora nearby pixel that is in
  hard shadow
• When a maximum search radius is reached then
  pixel is completely lit

16
Adding Outer Penumbra

• Project Search

17
Adding Outer Penumbra

• Assign attenuation value
• f rises up from 0 (no illum) to 1 (full illum)
• To do
• Define properties that affect size of penumbra
• Integrate those into the search scheme

hard shadow found
maximal extend
18
Adding Outer Penumbra

• Distance receiver/light

larger penumbra
19
Adding Outer Penumbra

• Distance receiver/light
• Varying rmax according to this distance

depth map search
20
Adding Outer Penumbra

• Distance receiver/occluder

larger penumbra
21
Adding Outer Penumbra

• Distance receiver/occluder
• Is known when a (possible) occluderis found
  (inside the rmax region)
• Dim f based on this distance
• Keep searching in rmax for minimum f

22
Adding Outer Penumbra

• Distance receiver/occluder

r
m
a
x
P

depth map search
23
Outer Penumbra Algorithm
search(P) rmax Pz fmin 1 search r
inside rmax found blocker at r ? rmax
rmax (Pz - depth_lightat r) f
clamp01(r/ rmax) fmin min(f, fmin ) return
fmin
receiver/light
receiver/occluder
24
Adding Inner Penumbra

• Previous slide
• Algorithm that adds a penumbra to the hard
  shadow (like Parkers method)
• Better
• Penumbra region that also smoothesthe inner
  region

25
Adding Inner Penumbra

• Search scheme can also be appliedto inner
  regions

c is user defined constant where inner and outer
meet (e.g. 0.5)
26
Self Shadowing Problem

• Problem when searchingwith constant Pz
• Surface can shadowitself

27
Self Shadowing Problem

• Solutions
• Divide scene into disjunct occludersand
  receivers
• Better additional search criterion
• Assign object IDs (like Hourcade)
• IDs should group logical, spatially
  relatedobjects
• Outer penumbra search id_light.. ! PID
• Inner penumbra search id_light.. PID

28
Self Shadowing Problem

• Object IDs
• May also fail if distinct objects are nearly
  adjacent
• No self-shadowing with soft shadows
• Use hard shadow for such cases
• No automatic ID assignment (yet)
• Simply used scene graph structure

29
Implementation

• Generating the input data
• Render depth id map fromlight source position
• e.g. store (ID mod 256) in stencil bufferand
  read back 24/8 packed depth_stencil
• Render depth id map fromcamera position

30
Implementation

• Shadow computation
• Render light camera id/depth map
• read back to host memory
• Perform search
• Result attenuation map(window resolution)
• Use attenuation map to modulate light source
  contribution

31
Implementation

• Subpixel Accuracy
• Keep fractional partwhen warping P to P
• Reduces quantizationartifacts

32
Implementation

• Adaptive Sampling
• Reduce number of searches
• Start with one search per NxN block
• Refine when
• Object IDs differ
• Hard shadow results differ
• If not, interpolate region
• Row / Column refinement

33
Implementation

• Example ID of 3 differs from rest

34
Results
35
Results
36
Results
37
Results
38
Results
39
Conclusions

• Good looking, soft penumbra regionsusing only
  information obtained froma single light source
  sample
• Independent of scene complexity
• like shadow mapping
• Runs at interactive frame rates
• Also suitable for ray tracing

40
Future Work

• Control of penumbra size
• Frustum settings / rmax affect size
• Better define area light in scene units
• Cases where object ID method fails
• Nearby objects
• Self-shadowing with soft shadows ?
• Speed-Up
• Distance map ?

41
Thank You !