Transcript Presentation - Computer Science Department, Technion

Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
Transitive Re-identification
Yulia Brand𝟏
Tamar Avraham𝟐
Michael Lindenbaum𝟐
𝟏Electrical Engineering Department
𝟐Computer Science Department
Technion - I.I.T. Haifa, Israel
This research was supported by the MAGNET program in the Israeli ministry of industry
and commerce, by the Israeli ministry of science and by the E. and J. Bishop research fund.
BMVC 2013
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
ReIDentification (ReID)
Camera A, time t
Camera B, time t+∆t
• upper image from: A. Bialkowski, S. Denman, P. Lucey, S. Sridharan, and C. B. Fookes. “A database for person re-identification in multicamera surveillance networks.” (DICTA 2012)
• lower images: courtesy of Marco Cristani
Introduction
Motivation
The Transitive Alg.
ReIDentification
Experiments
Summary & Future Work
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
ReIDentification
100
90
ICT – learning based
Recognition %
80
SDALF - direct
70
60
50
40
30
ICT trained on 3-7 data
SDALF for cameras 3-7
20
10
1
2
3
4
5
6
7
rank score
8
9
10
11
12
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
ReIDentification
Camera A
F
same
C
Camera B
F
not same
* Image from: L. Bazzani, M. Cristani, and V. Murino. Symmetry-driven
accumulation of local features for human characterization and reidentification.
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
Motivation
𝑪𝟏
𝑪𝟏
𝑪𝑵
𝑪𝟐
𝑪𝑵
𝑪𝟐
𝑪𝟓
𝑪𝟑
𝑪𝟓
𝑪𝟑
𝑪𝟒
𝑪𝟒
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
Motivation
By applying the
transitive algorithm:
𝑪𝟏
By recursively applying the
transitive algorithm:
𝑪𝟏
𝑪𝟏
𝑪𝑵
𝑪𝟐
𝑪𝑵
𝑪𝟐
𝑪𝑵
𝑪𝟐
𝑪𝟓
𝑪𝟑
𝑪𝟓
𝑪𝟑
𝑪𝟓
𝑪𝟑
𝑪𝟒
directly trainable pairs
𝑪𝟒
non-directly trainable pairs
𝑪𝟒
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
Motivation
𝑪𝟏
𝑪𝑵
𝑪𝟐
𝑪𝟓
𝑪𝟑
𝑪𝟒
𝑪𝑵+𝟏
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
The Transitive Algorithm
𝑪𝟏
𝑪𝑵
A
𝑪𝟏
𝑪𝟐
𝑪𝟓
𝑪𝟐
B
𝑪𝟑
C
𝑪𝟑
𝑪𝟒
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
The Transitive Algorithm
A
𝑺𝑨𝑩
B
𝑺𝑩𝑪
Training
A
B
C
Test
B
C
A
C
=
?
=
?
=
?
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
The Transitive Algorithm
A
B
100
90
ICT
Recognition %
80
C
70
SDALF
60
50
Naive ICT
40
30
ICT trained on 3-7 data
SDALF for cameras 3-7
ICT trained on 3-5 and 5-7 data
20
10
1
2
3
4
5
6
7
rank score
8
9
10
11
12
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
The Transitive Algorithm
𝒋
𝑺𝑨𝑩 = {(𝒙𝒊𝑨 , 𝒙𝑩 )}
𝒋
𝒀𝑨𝑩 (𝒙𝒊𝑨 , 𝒙𝑩 )
𝒋
𝑺𝑩𝑪 = {(𝒙𝒊𝑩 , 𝒙𝑪 )}
𝟏,
=
A
𝟎,
𝒊=𝒋
𝒋
𝒀𝑩𝑪 (𝒙𝒊𝑩 , 𝒙𝑪 )
𝒊≠𝒋
=
𝟏,
𝒊=𝒋
𝟎,
𝒊≠𝒋
𝑺𝑨𝑩
B
𝑺𝑩𝑪
C
𝑷 𝒀𝑨𝑪 𝒙𝑨 , 𝒙𝑪 =?
𝐏 𝒀𝑨𝑪 𝒙𝑨 , 𝒙𝑪 =
𝑷 𝒀𝑨𝑪 , 𝒀𝑨𝑩 = 𝒚𝑨𝑩 , 𝒀𝑩𝑪 = 𝒚𝑩𝑪 , 𝒙𝑩 𝒙𝑨 , 𝒙𝑪 ) 𝒅𝒙𝑩
𝒚𝑨𝑩 ∈{𝟎,𝟏} 𝒚𝑩𝑪 ∈{𝟎,𝟏} 𝒙𝑩 ∈𝑹𝒅
𝐏 𝒀𝑨𝑪 𝒙𝑨 , 𝒙𝑪 =
𝑷 𝒀𝑨𝑪 , 𝒀𝑨𝑩 = 𝒚𝑨𝑩 , 𝒀𝑩𝑪 = 𝒚𝑩𝑪 𝒙𝑨 , 𝒙𝑩 , 𝒙𝑪 )𝒇𝒙𝑩 (𝒙𝑩 ) 𝒅𝒙𝑩
𝒚𝑨𝑩 ∈{𝟎,𝟏} 𝒚𝑩𝑪 ∈{𝟎,𝟏} 𝒙𝑩 ∈𝑹𝒅
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
The Transitive Algorithm
𝒋
𝑺𝑨𝑩 = {(𝒙𝒊𝑨 , 𝒙𝑩 )}
𝒋
𝒀𝑨𝑩 (𝒙𝒊𝑨 , 𝒙𝑩 )
𝟏,
=
𝟎,
A
𝒋
𝑺𝑩𝑪 = {(𝒙𝒊𝑩 , 𝒙𝑪 )}
𝒊=𝒋
𝒋
𝒀𝑩𝑪 (𝒙𝒊𝑩 , 𝒙𝑪 )
𝒊≠𝒋
=
𝟏,
𝒊=𝒋
𝟎,
𝒊≠𝒋
B
C
𝑷 𝒀𝑨𝑪 𝒙𝑨 , 𝒙𝑪 =?
.
.
.
.
𝑷 𝒀𝑨𝑪 𝒙𝑨 , 𝒙𝑪 =
𝒙𝑩
𝟏−
𝑷 𝒀𝑨𝑩 𝒙𝑨 , 𝒙𝑩 𝑷 𝒀𝑩𝑪 𝒙𝑩 , 𝒙𝑪 𝒇𝑿𝑩 𝒙𝑩 𝒅𝒙𝑩
𝒙𝑩
𝑷 𝒀𝑨𝑩 𝒙𝑨 , 𝒙𝑩 𝑷 𝒀𝑩𝑪 𝒙𝑩 , 𝒙𝑪
𝒇𝑿𝑩 𝒙𝑩 𝒅𝒙𝑩
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
The Transitive Algorithm
A
B
𝑷 𝒀𝑨𝑪 𝒙𝑨 , 𝒙𝑪 =
𝒙𝑩
𝟏−
𝑷 𝒀𝑨𝑩 𝒙𝑨 , 𝒙𝑩 𝑷 𝒀𝑩𝑪 𝒙𝑩 , 𝒙𝑪 𝒇𝑿𝑩 𝒙𝑩 𝒅𝒙𝑩
𝒙𝑩
𝑷 𝒀𝑨𝑩 𝒙𝑨 , 𝒙𝑩 𝑷 𝒀𝑩𝑪 𝒙𝑩 , 𝒙𝑪
𝒇𝑿𝑩 𝒙𝑩 𝒅𝒙𝑩
C
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
The Transitive Algorithm
A
B
𝑷 𝒀𝑨𝑪 𝒙𝑨 , 𝒙𝑪 =
𝒙𝑩
𝟏−
𝑷 𝒀𝑨𝑩 𝒙𝑨 , 𝒙𝑩 𝑷 𝒀𝑩𝑪 𝒙𝑩 , 𝒙𝑪 𝒇𝑿𝑩 𝒙𝑩 𝒅𝒙𝑩
𝒙𝑩
𝑷 𝒀𝑨𝑩 𝒙𝑨 , 𝒙𝑩 𝑷 𝒀𝑩𝑪 𝒙𝑩 , 𝒙𝑪
𝒇𝑿𝑩 𝒙𝑩 𝒅𝒙𝑩
C
𝑷 𝒀𝑨𝑪 𝒙𝑨 , 𝒙𝑪 ≈
𝟏
𝑺𝑩
𝟏−
𝟏
𝑺𝑩
𝒙𝑩 ∈𝑺𝑩
𝑷 𝒀𝑨𝑩 𝒙𝑨 , 𝒙𝑩 𝑷 𝒀𝑩𝑪 𝒙𝑩 , 𝒙𝑪
𝒙𝑩 ∈𝑺𝑩
𝑷 𝒀𝑨𝑩 𝒙𝑨 , 𝒙𝑩 𝑷 𝒀𝑩𝑪 𝒙𝑩 , 𝒙𝑪
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
Synthetic Experiment 1
𝒙𝑩 × 𝒙𝑪
𝒙𝑨 × 𝒙𝑩
Recognition
Recognition
%%
𝒙𝑨 × 𝒙𝑪
-40 -20
-20
-40
-40 -20
100
100
100
00
0
20 40
40
20
20 40
00
0
-50
-50
-50
00
0
50
50
50
-50
-50
-50
50
50
50
00
000
0
20
20
20
𝒙𝑨 × 𝒙𝑪
40
40
40
𝒙𝑨 × 𝒙𝑪
60
60
60
-50
-50
-50
00
0
50
50
50
ICT trained
trained on
on AC
AC data
data
ICT
ICT
trained
Naive
ICT on AC data
Naive
ICT
Naive
TRID ICT
TRID
TRID
rank
rank
80
100 rank
80
100
score
score
80
100
score
𝒙𝑨 × 𝒙𝑪
𝑷 𝒀𝑨𝑪 𝒙𝑨 , 𝒙𝑪 :
𝑵𝒂𝒊𝒗𝒆 𝑰𝑪𝑻
B
50
50
50
50
50
50
00
0
-50
-50
-50
20
20
20
00
0
-20
-20
-20
A
𝑰𝑪𝑻 𝒕𝒓𝒂𝒊𝒏𝒆𝒅 𝒐𝒏 𝑨𝑪 𝒅𝒂𝒕𝒂
𝑻𝑹𝑰𝑫
C
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
Synthetic Experiment 2
A
𝒙𝑩 × 𝒙𝑪
𝒙𝑨 × 𝒙 𝑩
𝒙𝑨 × 𝒙𝑪
B
20
20
0
0
-20
-20
Recognition
% %
Recognition
-20
100
0
-20
0
20
50
50
50
0
50
0
0
-50
0
-50
-50
-50
-50
20
0
(d)
0
50
-50
50
C
-50
0
50
-50
0
50
(d)
100
50
50
0
0
0
0
10
20
30
40
50
60
70
10
20
30
40
𝒙𝑨 × 𝒙𝑪
50
60
70
𝒙𝑨 × 𝒙𝑪
ICT trained on AC data
Naive
ICT on AC data
ICT
trained
TRID ICT
Naive
80
90
100 rank
TRID
score
rank
80
90
100
score
𝒙𝑨 × 𝒙𝑪
𝑷 𝒀𝑨𝑪 𝒙𝑨 , 𝒙𝑪 :
𝑵𝒂𝒊𝒗𝒆 𝑰𝑪𝑻
𝑰𝑪𝑻 𝒕𝒓𝒂𝒊𝒏𝒆𝒅 𝒐𝒏 𝑨𝑪 𝒅𝒂𝒕𝒂
𝑻𝑹𝑰𝑫
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
SAIVT-SoftBio Experiment
• image from: A. Bialkowski, S. Denman, P. Lucey, S. Sridharan, and C. B. Fookes. “A database for person re-identification in multi-camera
surveillance networks.” (DICTA 2012)
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
SAIVT-SoftBio Experiment
[A B C] = C3 C5 C7
100
90
90
80
Recognition %
Recognition %
80
70
60
50
40
ICT trained on AC data
Naive ICT
TRID
SDALF
30
20
10
[A B C] = C1 C5 C7
100
0
2
4
6
8
10
70
60
50
ICT trained on AC data
Naive ICT
TRID
SDALF
40
30
20
10
12
0
2
4
6
8
[A B C] = C1 C5 C3
90
90
80
80
70
60
50
ICT trained on AC data
Naive ICT
TRID
SDALF
40
30
20
10
0
2
4
6
8
rank score
12
14
16
18
20
[A B C] = C1 C3 C8
100
Recognition %
Recognition %
100
10
rank score
rank score
10
12
70
60
50
ICT trained on AC data
Naive ICT
TRID
SDALF
40
30
20
14
10
0
5
10
rank score
15
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
SAIVT-SoftBio Experiment
[A B C] = C3 C5 C7
100
90
Recognition %
80
70
60
50
40
ICT trained on AC data
Naive ICT
TRID
SDALF
30
20
10
0
2
4
6
8
10
12
rank score
C
B
A
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
SAIVT-SoftBio Experiment
[A B C] = C3 C5 C7
100
90
90
80
Recognition %
Recognition %
80
70
60
50
40
ICT trained on AC data
Naive ICT
TRID
SDALF
30
20
10
[A B C] = C1 C5 C7
100
0
2
4
6
8
10
70
60
50
ICT trained on AC data
Naive ICT
TRID
SDALF
40
30
20
10
12
0
2
4
6
8
[A B C] = C1 C5 C3
90
90
80
80
70
60
50
ICT trained on AC data
Naive ICT
TRID
SDALF
40
30
20
10
0
2
4
6
8
rank score
12
14
16
18
20
[A B C] = C1 C3 C8
100
Recognition %
Recognition %
100
10
rank score
rank score
10
12
70
60
50
ICT trained on AC data
Naive ICT
TRID
SDALF
40
30
20
14
10
0
5
10
rank score
15
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
SAIVT-SoftBio Experiment
C
B
[A B C] = C1 C3 C8
100
90
80
Recognition %
A
70
60
50
ICT trained on AC data
Naive ICT
TRID
SDALF
40
30
20
10
0
5
10
rank score
15
Introduction
Motivation
Summary
The Transitive Alg.
Experiments
Summary & Future Work
Introduction
Motivation
The Transitive Alg.
Experiments
Summary & Future Work
Future Work
𝑪𝟏
𝑪𝟐
𝑪𝟑
𝑪𝟒
Introduction
Motivation
Thank-You.
The Transitive Alg.
Experiments
Summary & Future Work