: .
.4
3 :
1) , [35] [36] [37].
2) .
3) .
:
1) R-CNN [20],
2) Multi-Box [21],
3) Fast R-CNN [22].
. : GoogLeNet [23], Network in Network [24], AlexNet [8].
R-CNN
R-CNN Ross Girshick 2012 [20]. :
- Selective Search [25]
( )
N , (Support Vector Machine).
- , , .
.5. R-CNN
, - , , , 1 Selective Search 2 000 -.
Multi-Box
Multi-Box 2013 Google: Dumitru Erhan, Christian Szegedy, Alexander Toshev Dragomir Anguelov. [21] [26]
, Selective Search , - c ϵ [0,1], , - ( ).
GoogLeNet, [23].
, R-CNN, .
Fast R-CNN
2015 Ross Girshick [22]. , R-CNN. , R-CNN . , ( -), 1 [27].
|
|
Multi-Box . . Multi-Box c ϵ [0,1] , - , c ϵ [0,1] .
Fast RCNN PASCAL VOC 2012 R-CNN ~ 9x . 213 .
.6 Fast R-CNN
PASCAL VOC 2012 & PASCAL VOC 2007 20 .
ImageNet 2014. : , , .. , , , ImageNet & PASCAL VOC (200 20 ).
ImageNet 2014
ImageNet [28] [39]. 200 516840 . 60658 Flickr. , , . 200 .
1. ImageNet 2014
ImageNet2014 train | |
ImageNet2014 validation | |
ImageNet2014 test |
.7 ImageNet 2014
.8 ImageNet 2014
Pascal VOC 2012
Pascal VOC 2012 [29]. 20 . 22531 .
2. Pascal VOC 2012
PASCAL VOC 2012 train | |
PASCAL VOC 2012 validation | |
PASCAL VOC 2012 test |
.9 PASCAL VOC 2012
:
1) Caffe
2) C++
3) Python
4) Matlab R2014b
5) QT Creator
Caffe [30] open-source (BSD-2) . Caffe BVLC (Berkeley Vision and Learning Center), ++ 2013 , Python, atlab. Matlab.
Caffe :
1)
2)
3) CUDA
4)
Caffe :
CONVOLUTION . , ( ). , , .
|
|
. 10
POOLING c , .
. 11
INNER PRODUCT , .
. 12
ReLU Rectified-Linear Unit ( ).
:
TANH ( ).
:
SIGMOID .
:
EUCLIDEAN LOSS .
:
SIGMOID CROSS-ENTROPY LOSS
ACCURACY . .
SOFTMAX . [0,1].
LOCAL RESPONSE NORMALIZATION .
:
DROPOUT . , d ( 0.4, 0.5), INNER PRODUCT .
Caffe , , .
Caffe :
1) [31],
2) [32] [34],
3) [33].