Very Deep Residual Networks with Maxout for Plant Identification in - - PowerPoint PPT Presentation

SLIDE 1

Mi l a n Š u l c , Dmy t r

• Mi

s h k i n , J i ř í Ma t a s C e n t e r f

• r

Ma c h i n e P e r c e p t i

• n

D e p a r t m e n t

• f

C y b e r n e t i c s F a c u l t y

• f

E l e c t r i c a l E n g i n e e r i n g C z e c h T e c h n i c a l U n i v e r s i t y i n P r a g u e

Very Deep Residual Networks with Maxout for Plant Identification in the Wild

SLIDE 2

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 1

P l a n t Re c

• g

n i t i

• n

a t C MP b e f

• r

e C L E F ' 1 6 We w

• r

k e d

• n

n a r r

• w

e r p r

• b

l e m s w i t h h a n d

• c

r a f t e d f e a t u r e s w i t h s t a t e

• f
• t

h e

• a

r t r e s u l t s :

• B

a r k r e c

• g

n i t i

• n

: t e x t u r a l d e s c r i p t i

• n

[ 1 ]

• L

e a f r e c

• g

n i t i

• n

: d e s c r i b i n g t e x t u r e

• f

t h e l e a f i n t e r i

• r

a n d b

• r

d e r [ 2 ]

[1] Kernel-mapped histograms of multi-scale LBPs for tree bark recognition. Milan Šulc and Jiří Matas. IVCNZ 2013. [2] Fast features invariant to rotation and scale of texture. Milan Šulc and Jiří Matas. ECCV 2014, CVPPP workshop.

SLIDE 3

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 2

L e s s

• n

s f r

• m

p r e v i

• u

s P l a n t C L E F s

• B

e s t p e r f

• r

m i n g d e s c r i p t

• r

s :

• S

e p a r a t e n e t w

• r

k s f

• r

d i fg e r e n t c

• n

t e n t t y p e s d i d n ' t h e l p .

• S

i g n i fj c a n t e fg e c t

• f

b a g g i n g .

SLIDE 4

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 3

L e s s

• n

s f r

• m

C NN E v

• l

u t i

• n
• R

e s i d u a l N e t w

• r

k s [ 3 ] ( R e s N e t ) : B e s t r e s u l t s i n I L S V R C 2 1 5 a n d MS C O C O 2 1 5 .

• Ma

x

• u

t [ 4 ] a c t i v a t i

• n

f u n c t i

• n

l

• k

s p r

• m

i s s i n g , w h e n c

• m

b i n e d w i t h d r

• p
• u

t f

• r

b e t t e r r e g u l a r i z a t i

• n

.

[3] Deep Residual Learning for Image Recognition. Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun. CVPR 2016. [4] Maxout Networks. Ian J. Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron C. Courville, and Yoshua Bengio. ICML (3) 28 (2013): 1319-1327.

SLIDE 5

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 4

De e p Re s i d u a l Ne t w

• r

k s

• H

e e t a l . [ 3 ] s h

• w

e d t h a t r e s i d u a l c

• n

n e c t i

• n

s a c c e l e r a t e l e a r n i n g e v e n f

• r

e x t r e m e l y d e e p n e t w

• r

k s .

• We

b u i l d

• n

t h e R e s N e t

• 1

5 2 m

• d

e l p r e

• t

r a i n e d

• n

I m a g e N e t .

• 8

x d e e p e r t h a n V G G

• 1

9 [ 5 ] , b u t s t i l l l

• w

e r c

• m

p l e x i t y . V G G

• 1

9 : 1 9 . 6 b i l l i

• n

F L O P s . R e s N e t

• 1

5 2 :1 1 . 3 b i l l i

• n

F L O P s .

[3] Deep Residual Learning for Image Recognition. Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun. CVPR 2016. [5] Very deep convolutional networks for large-scale image recognition. Karen Simonyan and Andrew Zisserman. arXiv preprint arXiv:1409.1556 (2014).

SLIDE 6

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 5

T h e Ma x

• u

t

• Ma

x

• u

t [ 4 ] u n i t : ~ n e t w

• r

k a c t i v a t i

• n

f u n c t i

• n

.

• D

r

• p
• u

t i s p e r f

• r

m e d

• n

x , b e f

• r

e m u l t i p l i c a t i

• n

b y w e i g h t s .

[4] Maxout Networks. Ian J. Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron C. Courville, and Yoshua Bengio. ICML (3) 28 (2013): 1319-1327.

SLIDE 7

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 6

T h e Ma x

• u

t

• A

s i n g l e Ma x

• u

t [ 4 ] u n i t c a n b e i n t e r p r e t e d a s m a k i n g a p i e c e w i s e l i n e a r a p p r

• x

i m a t i

• n

t

• a

n y c

• n

v e x f u n c t i

• n

.

[4] Maxout Networks. Ian J. Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron C. Courville, and Yoshua Bengio. ICML (3) 28 (2013): 1319-1327.

SLIDE 8

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 7

C MP Ne t w

• r

k : Re s Ne t

• 1

5 2 w i t h Ma x

• u

t

• R

e s N e t

• 1

5 2 p r e

• t

r a i n e d

• n

I m a g e N e t .

• F

C l a y e r r e p l a c e d b y 4 p i e c e s

• f

F C l a y e r s , 5 1 2 n e u r

• n

s e a c h , f

• l

l

• w

e d w i t h Ma x

• u

t .

• D

r

• p
• u

t i s p e r f

• r

m e d

• n

t h e i n p u t s

• f

t h e F C l a y e r s .

• A

n

• t

h e r F C l a y e r i s a d d e d

• n

t h e t

• p

f

• r

c l a s s i fj c a t i

• n

.

SLIDE 9

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 8

P r e l i mi n a r y E x p e r i me n t s

• T

r a i n i n g s e t : 2 1 5 t r a i n i n g d a t a .

• V

a l i d a t i

• n

s e t :2 1 5 t e s t d a t a .

SLIDE 10

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 9

C MP S u b mi s s i

• n

s

• C

MP R u n 1 ( m a i n s u b m i s s i

• n

) :

• B

a g g i n g

• f

3 n e t w

• r

k s ( R e s N e t

• 1

5 2 + Ma x O u t ) .

• P

l a n t C L E F 2 1 6 t r a i n i n g s e t d i v i d e d i n t

• 3

f

• l

d s , e a c h n e t w

• r

k u s e s 2 f

• l

d s f

• r

t r a i n i n g .

• F

i n e

• t

u n i n g f

• r

1 1 K i t e r a t i

• n

s ( d u e t

• l

i m i t e d t i m e ) .

• C

MP R u n 2 :

• O

n l y

• n

e

• f

t h e t h r e e n e t w

• r

k s .

• C

MP R u n 3 :

• N

e t w

• r

k fj n e

• t

u n e d i n p r e l i m i n a r y e x p e r i m e n t s

• n

P l a n t C L E F 2 1 5 t r a i n i n g d a t a , 3 7 K i t e r a t i

• n

s .

SLIDE 11

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 10

Offjc i a l S c

• r

e : Me a n A v e r a g e P r e c i s i

• n

SLIDE 12

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 11

Re s u l t s u s i n g Me t a d a t a

• T

h e w i n n i n g s u b m i s s i

• n

( B l u e fj e l d ) :

• s

u m s a l l s c

• r

e s w i t h t h e s a m e O b s e r v a t i

• n

I D = t r a n s f

• r

m s t h e t a s k f r

• m

s i n g l e

• i

m a g e r e c

• g

n i t i

• n

t

• m

u l t i p l e

• i

m a g e r e c

• g

n i t i

• n

.

• Wh

a t i f

• t

h e r p i p e l i n e s u s e d O b s e r v a t i

• n

I D ?

• S

h

• u

l d r e c

• g

n i t i

• n

f r

• m

s i n g l e i m a g e a n d f r

• m

m u l t i p l e i m a g e s b e e v a l u a t e d s e p a r a t e l y ?

T e a m S i n g l e

• i

m a g e r e c

• g

n i t i

• n

[ m A P ] Mu l t i p l e

• i

m a g e s r e c

• g

n i t i

• n

[ m A P ] B l u e fj e l d R u n 2 / 4 6 1 . 1 % 7 4 . 2 % S a b a n c i U G e b z e T U R u n 1 7 3 . 8 % 7 9 . 3 % C MP R u n 1 7 1 . % 7 8 . 8 %

SLIDE 13

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 12

S u mma r y

• C

MP r e c

• g

n i t i

• n

s y s t e m :

• s

t a t e

• f
• t

h e

• a

r t R e s i d u a l N e t w

• r

k s ( R e s N e t

• 1

5 2 )

• a

d d e d a Ma x

• u

t l a y e r

• b

a g g i n g

• f

3 n e t w

• r

k s ( l i m i t e d t i m e )

• 3

r d

b e s t s c

• r

i n g t e a m i n t h e c h a l l e n g e , l

• k

i n g f

• r

w a r d t

• P

l a n t C L E F 2 1 7 !

• We

a r e i n t e r e s t e d i n

• t

h e r p l a n t r e c

• g

n i t i

• n

t a s k s .

• P

r e p a r i n g a n A n d r

• i

d a p p d e p l

• y

i n g a C N N

• n

t h e p h

• n

e .

SLIDE 14

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d . 13

Mo b i l e Ap p

• We

s t a r t e d d e v e l

• p

i n g a n A n d r

• i

d a p p d e p l

• y

i n g a d e e p C N N m

• d

e l

• n

t h e p h

• n

e .

• n
• n

e e d f

• r

I n t e r n e t c

• n

n e c t i

• n

:

• )

i n t h e fj e l d

• n
• t

a s

• u

r c e

• f
• b

s e r v a t i

• n

s :

• (
• A

s k f

• r

a d e m

• u

t d

• r

s .

SLIDE 15

C L E F 2 1 6 M. Š u l c , D . Mi s h k i n , J . Ma t a s : V e r y D e e p R e s i d u a l N e t w

• r

k s w i t h Ma x O u t f

• r

P l a n t I d e n t i fj c a t i

• n

i n t h e Wi l d .

T h a n k y

• u

!

Q u e s t i

• n

s

?

s u l c mi l a @c mp . f e l k . c v u t . c z

14