free-view watermarking for free view television alper koz, cevahir Çığla and a.aydın alatan
TRANSCRIPT
FREE-VIEW WATERMARKING FOR FREE VIEW TELEVISION
Alper Koz, Cevahir Çığla and A.Aydın Alatan
Outline
Introduction Proposed Watermarking Method Robustness Results for the case of known external calibration. Analysis on the watermark transformations during Light Field
Rendering Proposed solution for unknown external calibration Robustness results Summary and future works
Free View TV
Rendered Video
Cam. 0 Cam. 1(Arbitrary view)
A New Problem: Free-view Video Watermarking
Cam. 0 Cam. 1
METU EEE
METU EEE
METU EEE (?)
Arbitrary View
PROBLEM: How to embed the watermark such that the watermark can be extracted from a generated video for an arbitrary view?
Fundamental Case: One object static scenes
Rendered Image
Cam. 0 Cam. 1Imagery Camera
Light Field Rendering (LFR)
Light field:
Static scene
No occlusion
Lambertian Surfaces
Light Field Rendering
For each desired ray:
- Compute intersection with uv and st plane.
-Take closest ray
Variants:
-Bilinear in uv only
-Bilinear in st only
-Quadrilinear in (u,v,s,t)
Light Field Watermarking
W
W
W
W
Direct Approach: Embed the same watermark into each frame of lightfield slab.
uv plane
(Camera plane)
st plane
(focal plane)
),().,(.),(),( tsWtsHtsItsI uvuvuv
Iuv: Frame corresponds to the camera at (u,v).
Huv: Output image after high pass filtering Iuv.
: Watermark strength.
W: Watermark sequence generated from a Gaussian distribution with zero mean unit variance.
Light Field Watermarking
The watermark is inserted to the off-sheared images rather than embedding directly to camera frames.
Camera plane
Focal plane
Cam. 0 Cam. 2 Cam. 1
The same watermark is embedded to the pixels of different camera frames whose corresponding light rays intersectat the same point in the focal plane.
Camera plane
Focal plane
Cam. 1 Cam. 2
wo w1 w2 w3 w4
Watermark Detection
Rendered watermark
W Wren
Rendered Image
W
IrenHigh pass filtering
Normalized Correlation
Wren
Comparison to a threshold
1/0renI
renren
renren
WI
WI
ˆ
,ˆ
Robustness Results
32x32 camera 256x256 pixels 24 Bits RGB
Camera plane
(2,2,2)
(2,0,2)
(0,-2,2)
(-2,0,2)
(0,0,2)
1
-1
-1
1
z
x
y
Focal plane
Geometry of Buddha Light Field :
Robustness Results
Camera position=[0.5 0 2]; Image plane normal: [0 0 1]; focal length=2 (default);
Rendered image Rendered watermarked image
Robustness Results
Camera position =[2.7 2.7 0]; Rotation: [0 0 0]
Rendered image Rendered watermarked image
Transformation on Watermark Sequence in LFR
Bilinear Interpolation is utilized during LFR
Planar projective transformation between the original and rendered watermark.
Same transformation between imagery camera plane and focal plane.
Transformation on Watermark Sequence in LFR
How to find the projective planar transformation?
Find the points at which object surfaceand focal plane intersect.
Find the matches in the arbitraryview.
Determine the transformation between match pairs.
Transformation on Watermark Sequence in LFR
How to find the projective planar transformation?
Utilize two properties of these special points:
Light rays have same intensity Observed at similar coordinates in LF images
Properties are related with Light Fieldparametrization.
Transformation on Watermark Sequence in LFR
How to find the projective planar transformation?
Find feature points between rendered and all LF images.
Determine images having higher number of correspondences (four images).
Select feature points at same coordinates with similar intensity values.
Fit a planar transformation model between match pairs.
Transformation on Watermark Sequence in LFR
Transformation on Watermark Sequence in LFR
Apply the transformation to the original watermark.
Normalized correlation between the rendered watermark and the rendered image.
To handle small shidts in pixel locations :
utilize magnitude coefficients of the 2D Fourier transformation of the images
Robustness Results
Summary and future works
A novel problem is introduced, Free-view watermarking. Analyses on watermark transformation are examined.
Known external calibrations. Unknown external calibrations Bilinear interpolation in u-v plane.
Watermark is successfully detected from an aritrary view.
Extension of the algorithm for multiple depths and static scenes