Visual 3D Modeling from
Images
Introduction
u Goal
camera aroud
n Automatically extract a realistic 3D model by freely moving a
u Method
n Using a sequence of images
u Result
n 3D model: a scaled version of original object
Overall Flow Chart
Feature Points
u Principle of feature points extraction
feature point should be sufficiently different from neighborhoods after a
small dispacement: corner(角点)
n
n
u Feature points matching
7x7 pixel widow
n matching methods:
a. sum-of-square differences(SSD:最小方差和)
b. normalized cross-correlation(NCC:归一化互相关)
Figure 1:corresponding corners
Feature Points
u Problems
n
scaling, rotation , lighting conditions...
->Matching using affinely in variant regions
n
outliers
->Robust algorithm:RANSAC (RANdom SAmpling Consensus:随机抽样一致性算
法)
u Other approaches
n
n
lines or curves
SIFT(Scale-invariant feature transform)
Projective Geometry:Pinhole camera model
fX
fY
Z
f
f
PXx
P
diag
1
1
1
1
0
0
0
X
Y
Z
1
(
f
,
f
0|I)1,
P is the cmera projection matrix
Projective Geometry:two-view geometry
The foundametal matrix:
x’T Fx = 0
Foundamental matrix F can be computed with 7 or 8 points
Structure and motion
u Select two views
n
n
sufficient features are matched
the views should not be too close to each other so that the initial structure
is well-conditioned
u Initial frame
P = [I|0] and P’ = [R|t]