解读 Fast-RCNN（2）

解读 Fast-RCNN（2）

今天来分析Section 2 Fast R-CNN architecture and training

先来看一下作者对网络框架的叙述，

1, A Fast R-CNN network takes input as an entire image and a set of object proposals

2, The network first processes the whole image with several CONV and max pooling layers to

     produce a conv feature map

3, Then, for each object proposals a region of interest (RoI) pooling layer extracts a fixed-length vector

     from the feature map

4, Each feature vector is fed into a sequence of fuly connected layers that finally branch into two

     sibling output layers:

     one that produces softmax probability estimates over K object classes

                        plus a catch-all "background" class

     another layer that outputs four real-valued numbers for each of the K object class

     Each set of 4 values encodes refined bounding-box positions for one of the K classes      

恰如下图所示：          

作者对RoI pooling layer的叙述，

1, The RoI pooling layer uses max pooling to convert the feature inside any valid region of interest into

      a small feature map with fixed spatial extent of H×W where H and W are layer hyper-parameters that

     are independent of any particular RoI

2, an RoI is a rectangular window into a conv feature map

3, Each RoI is defined by a four-tuple (r,c,h,w)

4, RoI max pooling works by dividing the h×w RoI window into an H×W grid of sub-windows of approximate

     size h/H×w/W and then max-pooling the values in each sub-window into the corresponding output grid

     cell

5, The RoI layer is simply the special-case of the spatial pyramid pooling layers used in SPPnets in which

     there is only one pyramid level

总结一下，学习了这一段内容，了解到：

1，Fast R-CNN的整体架构

2，RoI pooling layers的作用以及思想源头，即SPPnets