adding cvpr13 poster3 2 - umiacsjhchoi/paper/cvpr2013_adding_poster.pdf · 2016-01-03 ·...
TRANSCRIPT
`
Adding Unlabeled Samples to Categories by Learned A6ributes h6p://umiacs.umd.edu/~jhchoi/addingbya6r
[1] Salakhutdinov, Torralba, Tenenbaum, “Learning to Share Visual Appearance for MulBclass Object DetecBon”, CVPR 2011 | [2] Rastegari et al., “ANribute Discovery via Predictable DiscriminaBve Binary Codes”, ECCV 2012 | [3] Malisiewicz, Gupta, Efros, “Ensemble of Exemplar-‐SVMs for Object DetecBon and Beyond”, ICCV 2011
Be6er classificaAon model
To obtain a be6er classifier
More training data
Jonghyun Choi Mohammad Rastegari Ali Farhadi Larry S. Davis
min
Ic2I,wvc ,w
ac
X
c
⇣↵J
vc (Ic, w
vc ) + �J
ac (Ic, w
ac )
⌘+M(I)
subject to
J
vc (Ic, w
vc ) = kwv
ck22 + �v
nX
i=1
⇠c,i
Ic,i · yc,i(wvcxi) � 1� ⇠c,i, 8i 2 {1, . . . , n}
J
ac (Ic, w
ac ) = kwa
c k22 + �a
nX
j=1
⇣c,j �nX
k=l+1
Ic,k
⇣w
ac�(xk)
⌘
Ic,j · yc,j(wac�(xj)) � 1� ⇣c,j , 8j 2 {1, . . . , n}
nX
k=l+1
Ic,k �, Ic,k = 1, 8k 2 {1, . . . , l}
M(I) =
XX
c1 6=c2
Ic1 · Ic2.• No human required • No underlying
distribution assumed
Max-‐margin Classifier on visual feature space
Max-‐margin Classifier on a6ribute space
Top-‐Lambda Selector
Mutual Exclusion
Joint opBmizaBon for discovering discriminaBve aNributes and unlabeled samples
Use$all$Except$each$
Score$high$
Given$Training$Samples$
• AcBve learning • Require human in the loop
• Semi-‐supervised learning • Unlabeled samples are
assumed to follow the same distribuAon of labeled samples
By learning data driven attributes based on [2]
There are many categories that have a few training samples
Category Name Init. NN ALC Cat. E+CMashed Potato 45.03 34.02 51.15 61.39 63.92Orange 29.84 16.29 26.97 40.61 41.05Lemon 32.21 27.58 32.43 35.37 34.23Green Onion 25.06 16.50 19.66 38.57 40.20Acorn 13.09 11.05 15.41 19.35 20.10Co↵ee bean 58.29 43.89 56.62 64.65 66.54Golden Retriever 14.54 15.57 12.61 17.54 18.61Yorkshire Terrier 29.62 13.62 27.63 41.41 45.65Greyhound 15.24 15.73 15.64 14.75 15.22Dalmatian 43.84 27.97 37.91 54.42 57.23Miniature Poodle 26.10 12.50 21.16 28.87 30.21Average 30.26 21.34 28.84 37.90 39.36
20 30 40 50 60 70
30
31
32
33
34
35
36
37
38
39
40
Number of Sample Added (!)
Me
an
Av
era
ge
Pre
cis
ion
(%
)
Init. Set+ by C Only+ by E+C
0 5 10 15 20 25 30 35 40 45 50 5520
25
30
35
40
45
50
55
Size of Initial Labeled Set
Me
an
Av
era
ge
Pre
cis
ion
(%
)
Init. Set
+ by C Only
+ by E+C
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Lab
el P
uri
ty
Mashed P
otato
Orange
Lemon
Green O
nion
Acorn
Coffee B
ean
Golden R
etriever
Yorkshire
Terri
er
Greyhound
Dalmatia
n
Min
iatu
re P
oodle
Average
+ by C only
+ by E+C
0
10
20
30
40
50
60
70
Av
era
ge
Pre
cis
ion
(%
)
Mas
hed P
otato
Ora
nge
Lemon
Gre
en O
nion
Acorn
Coffee
Bean
Gold
en R
etrie
ver
Yorksh
ire T
errie
r
Gre
yhound
Dalm
atia
n
Min
iatu
re P
oodle
Avera
ge
Init. SetE!SVMOurs
This research is partly supported by ONR MURI
Grant No. N00014-10-1-0934
Given labeled samples
A6ribute Mapper
Number of training samples in SUN 09 Dataset[1]
• ‘Init.’: iniAal labeled training set. • ‘NN’: addiAon by ‘nearest neighbor’ in visual feature space • ‘ALC’: addiAon by ‘acAve learning criteria (ALC)’ that finds the examples close to the current decision hyper-‐planes • ‘Cat.’: our method of select examples using categorical a6ributes only. • ‘E+C’: addiAon using both categorical and exemplar a6ributes.
Comparison with other methods
• Red: Using the iniAal labeled set (Init. Set) • Green: The augmented set by our method using categorical at-‐ tributes only (+ by C only) • Blue: Categorical+Exemplar a6ributes respecAvely. (+ by E+C)
Varying number of added images
Purity of added examples Categorical a6ributes only (+ by C only) Categorical+Exemplar a6ributes (+ by E+C).
Low Purity sAll Improves Accuracy
Size of IniAal Labeled Set
0
10
20
30
40
50
60
70
80
90
Mea
n A
vera
ge
Pre
cisi
on
(%
)
Mas
hed P
otato
Ora
nge
Lemon
Gre
en O
nion
Acorn
Coffee
Bean
Gold
en R
etrie
ver
Yorksh
ire T
errie
r
Gre
yhound
Dalm
atia
n
Min
iatu
re P
oodle
Avera
ge
Init. Set
+ Similar Only
+ 50 Similar Only
+ 50 by Our Method (C)
+ 50 GND Samples
The mAP gain for the smallest iniAal labeled set (5) is the highest as expected. When the number of samples is larger than 25, our method (+ by C only) does not improve the mAP much, although it sAll improves by 1.18 − 2.74%.
Our exemplar a6ribute discovery method (Sec. 4.2) to Exemplar SVM[3]
Our method outperforms the exemplar SVM in terms of category recogniAon accuracy by APs without the extra large negaAve example set (size = 50,000).
Retrieval rank change by the absence of a training sample
ej(xi) =µ
rg(xi)� ⌫
rj(xi)
More stable than Exemplar-‐SVM w/o large negaAve set • Non-‐convex: I is discrete, w’s are conBnuous
• Solve by block coordinate descent
• Navy: using the iniAal labeled set (baseline). • Blue: using only similar examples among the selected 50 examples. • Green: using 50 similar examples to compare with the result of our selected 50 examples
• Orange: including both similar and exact examples. • Red: using a set of 50 ground truth images, which is the best achievable accuracy (upper bound).
Training'Samples'
Build'A2ribute'Space'
Project'
Find'Useful'A2ributes'
Unlabeled'Samples'!!
!!
!!
!!Project'
Choose'Confident'Examples' Selected)by)Categorical)A+ributes)
Selected)by)Exemplar)A+ributes)
Ini=al)Labeled)Training))Examples)
(Full-‐set vs Leave-‐one-‐out Set)