3D Acquisition Using Shadows
3D Acquisition Using Shadows
3D Acquisition Using Shadows
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>3D</strong> <strong>Acquisition</strong> <strong>Using</strong> <strong>Shadows</strong><br />
Jean-Yves Bouguet<br />
and Pietro Perona<br />
California Institute of Technology<br />
http://www.vision.caltech.edu/bouguetj
Goal: <strong>3D</strong> reconstruction<br />
...<br />
<strong>3D</strong> model
State of the art<br />
• Accurate<br />
•Bulky<br />
• Complicated<br />
• Cost: >10k$
Weak structured lighting system
Desk<br />
Lamp<br />
Camera<br />
Stick or<br />
pencil<br />
Desk<br />
The idea<br />
Time t
The geometry<br />
Desk<br />
Lamp<br />
Camera<br />
p<br />
S<br />
Π<br />
P<br />
Stick or<br />
pencil<br />
O<br />
Camera<br />
S<br />
p<br />
Π<br />
Image<br />
P<br />
P<br />
Π d<br />
Stick<br />
= ( O,<br />
p)<br />
Desk<br />
∩Π
The geometry<br />
Desk<br />
Lamp<br />
Camera<br />
S<br />
λ<br />
Π<br />
Stick or<br />
pencil<br />
O<br />
Camera<br />
S<br />
Π<br />
λ<br />
Image<br />
Π d<br />
Stick<br />
Λ<br />
Π= ( S,<br />
Λ)<br />
Desk<br />
Λ = ( O, λ)<br />
∩Π<br />
d
The geometry S<br />
S<br />
λ 1<br />
λ 2<br />
Π<br />
O<br />
Camera<br />
Π<br />
λ 1<br />
λ 2<br />
Image<br />
Π d<br />
Λ 2<br />
Λ 1<br />
Π v<br />
Λ 1 = ( O, λ1)<br />
∩Π<br />
Λ = O, λ ) ∩Π<br />
2<br />
( 2<br />
d<br />
Π=<br />
( Λ1,<br />
Λ2)<br />
v
y<br />
Spatio-temporal<br />
processing<br />
x<br />
[Kanade’91,Curless’95]<br />
p<br />
t s(p) = 133.27<br />
Time t<br />
Temporal processing<br />
t s (p) = 133.27<br />
Time t<br />
Spatial processing<br />
x ref = 130.6<br />
Column pixel coordinate x
Camera calibration<br />
• Position of the desk plane<br />
• Internal parameters of the camera<br />
[Tsai’87]
S<br />
Δ<br />
t s<br />
Lamp calibration<br />
S∈Δ<br />
O<br />
Camera<br />
b<br />
S<br />
Δ<br />
t s<br />
h<br />
Image<br />
T<br />
B<br />
Pencil<br />
Π d<br />
T s<br />
Desk
Vertical plane<br />
calibration<br />
{ Π } d, λI<br />
λ Ι<br />
Πv<br />
O<br />
Camera<br />
λ Ι<br />
Image<br />
Π d<br />
Λ Ι<br />
Π v
Angel experiment<br />
Accuracy: 0.1mm over 10cm ~ 0.1% error
Skull experiment<br />
Accuracy: 0.1mm over 10cm<br />
~ 0.1% error
Textured objects
Other objects
Pot-pourri scan<br />
Accuracy: 0.5mm over 50cm ~ 0.1% error
Accuracy: 1mm over 50cm<br />
~ 0.5% error<br />
Scanning with the sun
Accuracy: 1cm over 2m<br />
~ 0.5% error<br />
Scanning with the sun
Real-time implementation<br />
• Performance:<br />
30Hz, 320x240, Pentium II 300MHz<br />
• Single shadow pass:<br />
20 - 30 seconds (600-900 frames)<br />
• Refined scanning: 1 - 2 minutes
Future work<br />
• Registration of multiple scans<br />
complete models<br />
• Extended light source model
If you are lazy...