m 15338 : depth map estimation software version 2
DESCRIPTION
M 15338 : Depth Map Estimation Software version 2. Olgierd Stankiewicz Krzysztof Wegner team supervisor: Marek Domański Chair of Multimedia Telecommunications and Microelectronics Poznan University of Technology, Poland. April, 27th 2008, Archamps. Outline. Depth map quality measurement - PowerPoint PPT PresentationTRANSCRIPT
1
M15338: Depth Map Estimation
Software version 2
April, 27th 2008, Archamps
Olgierd StankiewiczKrzysztof Wegner
team supervisor: Marek DomańskiChair of Multimedia Telecommunications and Microelectronics
Poznan University of Technology, Poland
2
Outline Depth map quality measurement
Ground-truth map View resynthesis
View synthesis tool Depth map estimation tools
Belief Propagation based estimation Accuracy refinement by mid-level
hypothesis Summary
3
Depth map quality
Commonly used: ‘Bad-Pixels’
Miss information about error magnitude and energy
Requires ground-truth disparity map
thresholdyxdyxG
thresholdyxdyxGyxpixelbad
),(),(0
),(),(1),(
4
Depth map quality NBP-SAD (Normalized Bad Pixel SAD)
NBP-SSD (Normalized Bad Pixel SSD)
Still, requires ground-truth disparity map
pixelsbadyx
yxdyxGpixelsbadofcount
SADNBP,
),(),(1
pixelsbadyx
yxdyxGpixelsbadofcount
SSDNBP,
2),(),(
1
5
Depth map quality measurement by view resynthesis
End-user never sees depth-map Resynthesis
No standarized method Tool employs straight-forward method
PSNR (Peak Signal-to-Noise Ratio)of resynthesized view as quality measure
6
Bad-Pixels vs PSNRBad-pixels vs PSNR of resynthesis
HSABM+OF [1]ThreeViewBP [9]
AdaptingBP [4]
SubPixDoubleBP [6]
AdaptOvrSegBP [7]
PlaneFitBP [8]
Our proposal BP
Double BP[5]
SSD+MF [3]
29
30
31
32
33
34
35
36
0,00% 1,00% 2,00% 3,00% 4,00% 5,00% 6,00% 7,00% 8,00% 9,00%
Bad-pixels [%]
PS
NR
[d
B]
7
View synthesis tool
Simple and straight-forward For linearly positioned stereo pairs
only Two disparity maps and
corresponding reference views Weighting of pixels from side-views,
translated according to their disparity
8
View synthesis tool
9
Belief Propagation based depth estimation tool
Alternative for Hierarchical-Shape Adaptive Block Matching
Employs message passing for optimization of disparity map
hierarchical processing in layers
Pixel differences (1-point SAD) used as observations
10
Message passing in Belief Propagation
mt s→d – message passed in t-th iteration from node s to node d, V(fp,fq) – cost of belief change from disparity fp to disparity fq.
11
Message in Belief Propagation
Single message contains information about all possible disparities
12
Hierarchical processing in Belief Propagation
Higher resolution
Lower resolution
from the lowest resolution to the full resolution in
coarse-to-fine manner
13
Belief propagation
Vpq(xp,xq) – transition cost in node q between disparity xp and xq
insisted by nodeł p
Vp(xp) – observation in node p about disparity xp (SAD value)mpq(xq) – message from node p to q about disparity xq
14
Belief propagation
Pot model Simpleand computationally efficient .
Stable beliefs are prefered
15
Belief propagation results
1th iteration 20 iterations 300 iterations
Middlebury test results – 1,65% of bad-pixelsBest Middlebury algorithm – 0,88% of bad-pixels
16
Bad-Pixels vs PSNRBad-pixels vs PSNR of resynthesis
HSABM+OF [1]ThreeViewBP [9]
AdaptingBP [4]
SubPixDoubleBP [6]
AdaptOvrSegBP [7]
PlaneFitBP [8]
Our proposal BP
Double BP[5]
SSD+MF [3]
29
30
31
32
33
34
35
36
0,00% 1,00% 2,00% 3,00% 4,00% 5,00% 6,00% 7,00% 8,00% 9,00%
Bad-pixels [%]
PS
NR
[d
B]
17
Accuracy refinement by mid-level hypothesis
Low computational cost Improves accuracy of disparity
map (number of disparity levels) Spatial resolution unchanged Focuses on unit-step edges in
disparity map
18
Unit-step edges
19
Mid-level hypothesis
Hypothesis spread along unit-step edges
20
Refinement by mid-level hypothesis
Pixel accurate disparity (1x) After refinement (8x)
21
Works over untextured regions
22
ResultsChange of bad-pixels relative to (x1) during iterative refinement
0,00%
20,00%
40,00%
60,00%
80,00%
100,00%
120,00%
x1 x2 x4 x8 x16
Bad
-pix
els
chan
ge
23
Summary
New version of experimental Depth Estimation software
Quality measurement problem with respect to multi-view applications
Simple view resynthesis tool Belief Propagation depth estimation
tool Novel technique for accuracy
refinement