You are on page 1of 1

Maximum Margin Metric Learning Over Discriminative Nullspace for Person Re-identification

T M Feroz Ali and Subhasis Chaudhuri


Vision and Image Processing Lab, Indian Institute of Technology Bombay, India
{ferozalitm, sc}@ee.iitb.ac.in

Maximum Margin projection and


CUHK01 dataset (single-shot) VIPeR dataset
Person Re-identification kernel mapping
Methods R-1 R-10 R-20 Methods R-1 R-10 R-20
Step 1 Step 2
• Matching pedestrians over multiple cameras MLFL 34.30 65.00 75.00 PolyMap 36.8 83.7 91.7

• Non-overlapping cameras LOMO+XQDA 50.00 83.40 89.51 LOMO+XQDA 40.0 80.5 91.1

• Unseen persons: Testing identities unseen while KNFST 52.80 84.97 91.07 QALF 30.2 62.4 73.8
training CAMEL 57.30 -- -- MLAPG 40.7 82.3 92.4
GOG+XQDA 57.89 86.25 92.14
KNFST 42.3 82.9 92.1
WARCA 58.34 -- -- SSSVM 42.1 84.3 91.9
? NK3ML 67.09 91.85 95.92
GOG+XQDA 49.7 88.7 94.5
TPC 53.70 91.00 96.30
SCSP 53.5 91.5 96.7
Probe Quadruplet 62.55 89.71 --
l1-graph 41.5 -- --
Gallery Semantic 32.70 64.40 76.30
NK3ML 99.8 100 100
M.Ensembles 53.40 84.40 90.50
DGD 38.6 -- --
Nullspace Kernel Maximum Margin Metric Normalized Kernel Maximum Margin Criterion DLPAR 72.30 94.90 97.20
SCNN 37.8 66.9 --

Learning (NK3ML) • Kernalization r CUKH01 dataset (multi-shot)


Shi et al. 40.9 -- --

maximize   (M  N)
k k
Methods R-1 R-10 R-20
S-LSTM 42.4 79.4 --
Challenges Null Foley Sammon Transform
αk
k 1 TPC 47.8 84.8 91.1
• Normalization s.t.  k K k  1,
T
k  1,..., r
l1-Graph 50.10 -- --
SSM 53.7 91.5 96.1
LOMO+XQDA 61.98 89.30 93.62
• Large variations in scene charecteristics constraints Spindle 53.8 83.2 92.1
• Fisher criterion: w SB w T CAMEL 62.70 -- --
MuDeep 43.0 85.8 --
Ranked results
-- pose J F ( w)  T MLAPG 64.24 90.84 94.92
-- viewpoint w SW w • Solution: (M  N ) k  k K k SSSVM 65.97 -- --
OLMANS 45.0 85.0 93.6

-- illumination KNFST 66.07 91.56 95.64 DLPAR 48.7 85.1 93.0

-- camera color variations maximize {J F ( wi )} j  1, . . . , i  1 GOG+XQDA 67.28 91.77 95.93 PDC 51.3 84.2 91.5
||wi || 1, wi w j  0
T

-- background NK3ML 76.77 95.58 98.02 SHAPE 62.0 -- --


NK3ML algorithm DGD 66.60 -- -- Semantic 41.6 86.2 95.1

• Low resolution • Discriminative Nullspace T OLMANS 68.44 92.67 95.88 DCIA 63.9 87.5 --

• Disjoint training and w S B w  0,


T
w SW w  0 • Step 1: Map to low dimensional nullspace. SHaPE 76.00 -- --

testing identities • Step 2: Map to the high dimensional kernel space Spindle 79.90 97.10 98.60 deep learning methods
S B : between class scatter
• Small sample size problem • Solution:
w  ( Z  T  ZW ) SW : within class scatter • Advantages: Another Application: Object Verification
ST : total class scatter -- Least possible
Z T : nullspace of ST within class variance
General Approaches -- Maximum
ZW : nullspace of SW
margin between classes
• Feature descriptor design -- Closed form
• Metric Learning solution
• Deep learning Experiments
-- No free
• Re-ranking parameters to tune
GRID dataset
PRID450S dataset
Contributions Methods R-1 R-10 R-20 Methods - R-1 R-10 R-20
MtMCML 14.08 45.84 59.84 WARCA 24.58 -- --
KNFST 14.88 41.28 50.88
• Address small sample size (SSS) problem in re-ID. SCNCD 41.60 79.40 87.80
LOMO+XQDA 16.56 41.84 52.40
• Propose Nullspace Kernel Maximum Margin Metric PolyMap 16.30 46.00 57.60
CSL 44.40 82.20 89.80
• Disjoint training and test identities
Learning (NK3ML). Maximum Margin Criterion TMA 52.89 85.78 93.33
MLAPG 16.64 41.20 52.96 • Small sample size (SSS) problem
KNFST 59.47 91.96 96.53
• Inter-class margin: KEPLER 18.40 50.24 61.44
-- Collapse intra-class LOMO+XQDA 59.78 90.09 95.29
samples to singular points d (Ci , C j )  d (mi , m j )  s(Ci )  s(C j )
DR-KISS 20.60 51.40 62.60 Conclusions
SSSVM 22.40 51.28 61.20 SSSVM 60.49 88.58 93.60
-- Maximize inter-
class margin
1 c c
J   pi p j d (Ci , C j )
SCSP 24.24 54.08 65.20 GOG+XQDA 68.00 94.36 97.64 • Efficiently address small smaple size (SSS) problem
-- Use kernel mapping 2 i 1 j 1 GOG+XQDA 24.80 58.40 68.88 NK3ML 73.42 96.31 98.58 • NK3ML
for better separability NK3ML 27.20 60.96 71.04 Semantic 44.90 77.50 86.70 – collapses within-class variance to zero
SSDAL 22.40 48.00 58.40 SSM 72.98 96.76 99.11 – maximum margin between classes
• Optimization: r

• State-of-the-art-results on four datasets. maximize


vk
 v (S
k 1
k B  SW )vk SSM 27.20 61.12 70.56 • State-of-the-art results
• 99.8 % rank-1 accuracy on VIPeR dataset.
OL-MANS 30.16 49.20 59.36 pre/post-processing methods – 99.8 % rank-1 on VIPeR dataset
vk vk  1, k  1,..., r
T
subject to
• Applicable on other problem domains. • Applicable for other object verification problems
RESEARCH POSTER PRESENTATION DESIGN © 2012
RESEARCH POSTER PRESENTATION DESIGN © 2015
www.PosterPresentations.com
www.PosterPresentations.com

You might also like