Soccer Broadcast Video Summarization - CRCV

Soccer
Broadcast
Video Summarization
Soomro,
Smith,
Weigandt
Summarization
Papers
Action
Occurrences
Causality
Matrix
Directed
Acyclic Graph
Currently
June 19, 2013
Soccer Broadcast Video
Summarization
Khurram Soomro
Levi Smith
Christian Weigandt
Faculty Mentors:
Dr. Mubarak Shah
Dr. Niels Lobo
University of Central Florida
Center for Research in Computer Vision
crcv.ucf.edu
UCF
1/10
Shmuel Peleg’s Papers
Soccer
Broadcast
Video Summarization
Soomro,
Smith,
Weigandt
Creates compact summarization of events, moving
temporally-distinct actions into the same time-space
Use single stationary cameras
Summarization
Papers
Group activities do not get well-represented by the synopsis
Action
Occurrences
Disregards temporal ordering of actions in summary
Causality
Matrix
Not good for edited videos
Directed
Acyclic Graph
Currently
June 19, 2013
UCF
2/10
Shmuel Peleg’s Papers
Soccer
Broadcast
Video Summarization
Soomro,
Smith,
Weigandt
Summarization
Papers
Action
Occurrences
Causality
Matrix
Directed
Acyclic Graph
Currently
June 19, 2013
UCF
3/10
Other Sports’ Summarization Papers
Soccer
Broadcast
Video Summarization
1
Soomro,
Smith,
Weigandt
Summarization
Papers
Not robust to different broadcasting styles
2
Causality
Matrix
Currently
June 19, 2013
Summarize by detecting scenes which contain ’plays’
Not applicable to sports with continuous play such as
soccer or hockey
Action
Occurrences
Directed
Acyclic Graph
Summarize by detecting scene cuts/transitions and
appearance of certain features
3
Features include information such as camera angle or
players on the field
Do not learn actions. Only learn to extract important
frames.
UCF
4/10
Action Occurrences
Soccer
Broadcast
Video Summarization
Soomro,
Smith,
Weigandt
Summarization
Papers
Action
Occurrences
Causality
Matrix
Directed
Acyclic Graph
Currently
June 19, 2013
Action
Arguing
Attempt At Goal
Card Red Yellow
Celebrating
Corner
Foul
Free Kick
Goal
Goal Kick
Injury
Offside
Penalty
Substitution
Throw In
1
4
47
5
5
13
36
16
11
17
6
7
0
4
31
2
0
39
1
20
7
20
9
36
3
2
11
0
6
17
3
1
25
5
10
11
21
12
21
4
15
10
0
4
24
4
3
44
1
13
11
29
17
24
5
5
13
0
4
29
UCF
5
6
31
5
12
7
59
25
26
9
15
12
0
5
20
6
3
38
0
9
11
17
9
12
6
7
0
0
4
16
7
5
43
7
7
12
42
21
16
7
6
7
0
3
34
8
4
35
0
13
15
28
9
9
6
2
4
0
6
35
9
9
49
4
10
10
36
21
15
9
3
6
0
6
23
Totals
35
351
28
99
97
288
139
170
66
61
70
0
42
229
1675
5/10
Causality Matrix
Soccer
Broadcast
Video Summarization
Soomro,
Smith,
Weigandt
p(A|B) =
Summarization
Papers
Action
Occurrences
Causality
Matrix
Directed
Acyclic Graph
Currently
June 19, 2013
|A ∩ B|
|B|
|A ∩ B| is defined as the number of times where A is the
first event that occurs within t seconds of B.
|B| is defined as the number of times that action B
occurred.
UCF
6/10
Causality Matrix
Soccer
Broadcast
Video Summarization
Soomro,
Smith,
Weigandt
Summarization
Papers
Action
Occurrences
Causality
Matrix
Directed
Acyclic Graph
Currently
June 19, 2013
UCF
7/10
Directed Acyclic Graph
Soccer
Broadcast
Video Summarization
Soomro,
Smith,
Weigandt
Each node will be the predicted probability using SVM.
Use causality matrix as weights on edges.
Summarization
Papers
Action
Occurrences
Causality
Matrix
Directed
Acyclic Graph
Currently
June 19, 2013
...
...
UCF
8/10
Directed Acyclic Graph
Soccer
Broadcast
Video Summarization
Soomro,
Smith,
Weigandt
Use iterative approximation approach to find the path with
max flow
Summarization
Papers
Action
Occurrences
After determining labels, we will use graph clustering to
find important segments of the game
Causality
Matrix
Directed
Acyclic Graph
Use ground truth of highlights to learn significance priors
for each event
Currently
June 19, 2013
UCF
9/10
Currently
Soccer
Broadcast
Video Summarization
Soomro,
Smith,
Weigandt
Summarization
Papers
Action
Occurrences
Causality
Matrix
Directed
Acyclic Graph
Calculated STIP and MBH features for each of the action
clips
Generating codebook for bag of words
Look into other possible approaches such as HMM, CRF,
etc.
Currently
June 19, 2013
UCF
10/10