Slide 6
Slide 6 text
w $MBTTBDUJWBUJPONBQQJOH<;IPV
$713`>
)JHIMJHIUTNPTUJOGPSNBUJWFJNBHFSFHJPOTSFMFWBOUUPQSFEJDUFEDMBTT
$PNQVUFTXFJHIUFETVNPGGFBUVSFNBQTPGMBTUDPOWPMVUJPOBMMBZFSUPPCUBJODMBTTBDUJWBUJPO
NBQT
"UUFOUJPONBQGPSJNBHFDMBTTJ
fi
DBUJPO
$POW@
$POW@
$POW@
$POW@
$POW@
$POW@
$POW@
$POW@
$POW@
$POW@
$POW@
$POW@
$POW@
(MPCBMBWFSBHFQPPMJOH
0VUQVU
ʜ
"WFSBHF
"WFSBHF
"WFSBHF
'FBUVSFNBQ
w1 w2 w3
º º º
ʴ ʴ ʴ ɾɾɾ ʹ
$MBTTBDUJWBUJPONBQQJOH
"UUFOUJPONBQ
ʜ
ʜ
ʜ
"VTUSBMJBOUFSSJFS
w1
w2
w3
w1 *
+ w2 *
+ … + wn *
Class
Activation
Map
(Australian terrier)
=
Class Activation Mapping
Figure 2. Class Activation Mapping: the predicted class score is mapped back to the previous convolutional layer to generate the class
activation maps (CAMs). The CAM highlights the class-specific discriminative regions.
Here we ignore the bias term: we explicitly set the input
bias of the softmax to 0 as it has little to no impact on the
classification performance.
By plugging Fk
=
x,y
fk
(x, y) into the class score,
Sc, we obtain
Sc
=
k
wc
k
x,y
fk
(x, y) =
x,y k
wc
k
fk
(x, y). (1)
We define Mc as the class activation map for class c, where
each spatial element is given by
Mc
(x, y) =
k
wc
k
fk
(x, y). (2)
Thus, Sc
=
x,y
Mc
(x, y), and hence Mc
(x, y) directly
indicates the importance of the activation at spatial grid
(x, y) leading to the classification of an image to class c.
Intuitively, based on prior works [34, 30], we expect each
unit to be activated by some visual pattern within its recep-
Figure 3. The CAMs of two classes from ILSVRC [21]. The maps
highlight the discriminative image regions used for image classifi-
cation, the head of the animal for briard and the plates in barbell.