SIGGRAPH 2010 Talk Slides (PDF)

Using blur to 

affect perceived 

distance and size 

Robert Held 1,2 

Emily Cooper 1 

James O’Brien 1 

Martin Banks 1 

1 

UC Berkeley 

2 

UC San Francisco 

SIGGRAPH 2010 Los Angeles, CA 

Sunday, August 1, 2010

Sunday, August 1, 2010 

2


3


4

Blur in cinema 

• Minimize blur 

• Result: Scale models appear life-sized 

Images copyright Lucasfilm Ltd. 

5 


Goals 

1. Review optics of blur 

2. Determine how blur acts as a distance 

(and size) cue 

3. Develop tips and rules for changing blur 

6 


Optics of blur 

A 

Imaging Plane 

Focal Distance 

z0 

s0 

7 



A 

Imaging Plane 


z0 

s0 

7 



A 

Imaging Plane 

c 

Target Distance 

z1 

s1 


z0 

s0 

8 



A 

Imaging Plane 

c 


z1 

s1 


z0 

s0 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

8 


Blur in cinema, revisited 

• Minimize blur (small aperture, long exposure) 

• Result: Scale models appear life-sized 

Images copyright Lucasfilm Ltd. 

9 


A = 4.5mm 

A = 60m 

10 


“Fake blur” (tilt-and-shift lens) 

Lens 

Imaging Plane 

Focal Plane 

11 



Imaging Plane 

Focal Plane 

Lens 

12 



Imaging Plane 

Focal Plane 

Lens 

12 



Imaging Plane 

Focal Plane 

Lens 

12 


Information from blur 

A 

Imaging Plane 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 


z1 

s1 


z0 

s0 

13 



A 

Imaging Plane 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 


z1 

s1 


z0 

s0 

Typical values for human eye: 

A = Pupil size (~4.6mm) 

s0 = Eye length (~17mm) 

13 



A 

Imaging Plane 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 


z1 

s1 


z0 

s0 

Typical values for human eye: 

A = Pupil size (~4.6mm) 

s0 = Eye length (~17mm) 

Important terms: 

Blur magnitude: c 

Focal (absolute) distance: z0 

Relative distance: z1/z0 

13 



14 



c 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

15 



c 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

z 0 = A 

s0 

c 

1 − z 0 

z 1 

 

15 



• Blur alone cannot reveal 

absolute distance 

Absolute Distance (z0) z0 (m) 

100 

10 

1.0 

0.1 

0.01 

0.001 

4.6mm 

z 0 = A 

pupil diameter: 

mean = 4.6mm 

s.d. = 1.0mm 

0.6 

0.60mm 

s0 

c 

17mm 

1 − z 0 

z 1 

 

retinal blur = 2° 

1 2 

Relative Distance (z 1 /z 0 ) 

z1/z0 

16 



• Blur alone cannot reveal 

absolute distance 


100 

10 

1.0 

0.1 

0.01 

0.001 

4.6 ± 1.0mm 

z 0 = A 


mean = 4.6mm 

s.d. = 1.0mm 

0.6 

0.60mm 

s0 

c 

17mm 

1 − z 0 

z 1 

 

retinal blur = 2° 

1 2 


z1/z0 

17 


Other information 

• Perspective information 

can reveal z1/z0 

z 0 = A 

s0 

c 

1 − z 0 

z 1 

 


100 

10 

1.0 

0.1 

0.01 

0.001 

0.6 

1 2 


z1/z0 

18 


Model 

• Combined with relative depth information, 

blur can act as a cue to absolute distance 

• Bayesian approach: 

Absolute Distance (z0) (m) 

z0 

100 

10 

1.0 

0.1 

0.01 

0.001 

Depth-from-blur Likelihood 


mean = 4.6mm 

s.d. = 1.0mm 

0.6 

retinal blur = 0.1° 

1 2 

Depth-from-perspective Likelihood 

X = 

Combined Depth Estimate 

0.6 1 2 0.6 1 2 


z 1 /z 0 

19 


Recovering absolute distance 


z0 

100 

10 

1.0 

0.1 

0.01 

0.001 

Retinal Blur = 0.60mm 

0.6 

1 2 

Relative Distance (z(d) 

1 /z 0 ) 

z 1 /z 0 


100 

10 

1.0 

0.1 

0.01 

0.001 

0.6 

1 2 

Relative Distance (z 1 /z 0 ) z1/z0 

20 




z0 

100 

10 

1.0 

0.1 

0.01 

0.001 


0.6 

1 2 


1 /z 0 ) 

z 1 /z 0 


100 

10 

1.0 

0.1 

0.01 

0.001 

0.6 

1 2 


21 




z0 

100 

10 

1.0 

0.1 

0.01 

0.001 


0.6 

1 2 


1 /z 0 ) 

z 1 /z 0 


100 

10 

1.0 

0.1 

0.01 

0.001 

0.6 

1 2 


22 



Absolute Distance (z z0 0 ) (m) 

100 

10 

1.0 

0.1 

0.01 

0.001 

Predicted perceived 

distance: 8cm 

0.6 

1 2 


z1/z0 

23 


Approximating blur 

Consistent blur Aligned blur gradient 24 


Aligned blur gradient 

Absolute Distance (z 0 ) (m) 

z0 

100 

10 

1.0 

0.1 

0.01 

0.001 

0.6 

1 2 


Abs. Distance Distribution 

z 1 /z 0 

25 

• Predicted perceived distance: ~8cm 

• Expect weaker influence of blur due to variance 


Approximating blur (very badly) 

Consistent blur Unaligned blur gradient 26 


Unaligned blur gradient 

Absolute Distance (z 0 ) (m) 

z0 

100 

10 

1.0 

0.1 

0.01 

0.001 

0.6 

1 2 


z1/z0 

Abs. Distance Distribution 

• Predicted perceived distance: ambiguous 

• Expect weakest miniaturization effect, if any 

27 


Experiment 

• 7 sample scenes from GoogleEarth 

• 

unaligned gradient blur 

• 5 blur magnitudes 

Each scene rendered sharply and with consistent, aligned gradient, and 

Enter camera 

distance in 

meters: 

0.010 

Fixation point (0.5s) Stimulus (3s) Response 

28 


Results 

Reported Distance 

(Normalized) 

1.0 

Subject RRR 

0.1 

0.01 

0.001 

0 0.1 0.2 0.3 0.4 0.5 

All Subjects 

0 0.1 0.2 0.3 0.4 0.5 

Simulated Focal Distance (m) 

blur condition: 

consistent aligned gradient unaligned gradient 

29 


Semi-automated Algorithm 

30 



31 



31 


Choosing A for desired depth of field 

A 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

z1 

s1 

z0 

s0 

32 



A 

b 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

z1 

s1 

z0 

s0 

32 



A 

b 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

z1 

s1 

z0 

s0 

Blur in angular units: c 

b = 2 arctan 

2s 0 

≈ c 

s 0 

32 



A 

b 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

z1 

s1 

z0 

s0 


b = 2 arctan 

2s 0 

≈ c 

s 0 

A 

z 0 

 

1 − z 0 

b = 

z 1 

 

32 



A 

b 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

z1 

s1 

z0 

s0 


b = 2 arctan 

2s 0 

≈ c 

s 0 

• Blur depends on scene structure only 

• For natural blur, set camera aperture 

to pupil size (~4.5mm) 

b = 

A 

z 0 

 

1 − z 0 

z 1 

 

32 


Connecting cues: Blur and disparity 


z1 

Disparity of target 

Fixation Distance 

relative to fixation: 

z0 

s0 

= I 

δ = X L − X R 

33 

 

s0 

z 0 

 

1 − z 

0 

z 1 

I 

XL 

XR 



Disparity of target relative to fixation: 

δ = I 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

Diameter of blur circle: 

c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

34 




δ = I 

 

s0 

z 0 

 

1 − z 0 

z 1 

 


c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

c =(A/I)|δ| 

• In natural viewing, blur is proportional to disparity 

34 




δ = I 

 

s0 

z 0 

 

1 − z 0 

z 1 

 


c = A 

 

s0 

z 0 

 

1 − z 0 

z 1 

 

c =(A/I)|δ| 

c ≈ |δ| 

12 

• In natural viewing, blur is proportional to disparity 

• Practical application: Natural stereo content should be generated 

with camera apertures ~1/12 the camera baseline 

34 


Discussion 

• Blur is deeply connected to 

distance 

• Also closely related to other 

distance cues 

• Modeling and understanding 

the relationship between blur 

and other depth information 

helps us understand how to 

make blur appear natural 

35 


Discussion 

• Once we know how to make 

blur look natural, we can 

intentionally modify to create 

perceptual effects 

• Tilt-shift, model photography 

were gross modifications 

• Blur-based effects in stereo 

photography deserve attention 

36 


Acknowledgments 

The authors thank the following people for their valuable input: 

• Björn Vlaskamp 

• Johannes Burge 

• Kurt Akeley 

Funding: 

• NSF Graduate Research Fellowship 

• NDSEG Graduate Research Fellowship 

• NIH R01 EY012851 

• NSF BCS-0117701 

Original city images and data from GoogleEarth are copyright Terrametrics, 

SanBorn, and Google. 

37 


Tilt-shift effect 

38 


Approximating blur 

Consistent blur 

Aligned 

blur gradient 

% Difference 

(blur-circle diameter): 

−1000 −100 −10 0 10 100 1000 

39

SIGGRAPH 2010 Talk Slides (PDF)

Create successful ePaper yourself

Delete template?

Save as template?