Transcript Slide 1

Learning the “Epitome”
of a Video Sequence
Information Processing Workshop 2004
Vincent Cheung
Probabilistic and Statistical Inference Group
Electrical & Computer Engineering
University of Toronto
Toronto, Ontario, Canada
Advisor: Dr. Brendan J. Frey
Aug. 11, 2004
University
of Toronto
Information Processing Workshop 2004
Outline
● Image epitome
►
►
What?
Why?
● Implementation computation issues
►
Efficiently implementing the learning algorithm
● Video epitome
►
►
1 / 12
Extension to videos
Video inpainting
Cheung
Learning
Image
Information Processing Workshop 2004
Image Epitome
● Jojic, N., Frey, B., & Kannan, A. (2003). Epitomic
analysis of appearance and shape. In Proc. IEEE
ICCV.
Video
● Miniature, condensed version of the image
● Accurately accounts for the interesting properties
of the image
● Applications
►
►
►
►
2 / 12
object detection
texture segmentation
image retrieval
compression
Cheung
Image Epitome Examples
Video
Learning
Image
Information Processing Workshop 2004
3 / 12
Cheung
Learning the Image Epitome
Learning
Image
Information Processing Workshop 2004
Training Set
Video
Input image
Bayesian
network
T1
Unsupervised
eLearning
– epitome
T2
Z1
4 / 12
Epitome
e
Sample
Patches
Z2
…
TM
Tk – mapping
Zk – image patch
ZM
Cheung
Shifted Cumulative Sum Algorithm
Video
Learning
Image
Information Processing Workshop 2004
(1, 1), (i, j)
– row 1
+ row (P+1)
– col 1
+ col (P+1)
X
Cumsum
(2, 1), (i+1, j)
5 / 12
–
N
N
+
K
e (1, 2), (i, j+1)
2
(i,j)
K
+
+
-
-
+ pixel (1,1)
+ pixel
(P+1, P+1)
+N
(2, 2), (i+1,
N j+1)
Cheung
Collecting Sufficient Statistics
Video
Learning
Image
Information Processing Workshop 2004
X
P
e
Ta
P
P
Tß
6 / 12
P
Cheung
Video
Learning
Image
Information Processing Workshop 2004
Extending Epitomes to Videos
● Desire a miniature, condensed version of a video
sequence
● Want it to accurately account for the interesting
properties of the video
● Applications
►
►
►
►
►
►
►
7 / 12
optic flow
segmentation
texture transfer
layer separation
compression
noise reduction
inpainting
Cheung
Video Epitome
Learning
Image
Information Processing Workshop 2004
Training Set
Video
Input Video
Sample
Patches
Unsupervised
Learning
Video
Epitome
Frame 3
Frame 2
Frame 1
8 / 12
Cheung
Video Epitome Example
Video
Learning
Image
Information Processing Workshop 2004
Spatially Compressed
Temporally Compressed
9 / 12
Cheung
Learning
Image
Information Processing Workshop 2004
Video Inpainting (1)
● Fill in missing portions of a video
►
►
damaged films
occluding objects
Video
● Reconstruct the missing pixels from the video epitome
10 / 12
Cheung
Video Inpainting (2)
Video
Learning
Image
Information Processing Workshop 2004
11 / 12
Cheung
Information Processing Workshop 2004
Conclusion
● Improved the efficiency of learning image
epitomes
● Extended the concept of epitomes to video
sequences
● Demonstrated the ability of video epitomes to
model motion patterns through video inpainting
12 / 12
Cheung