ROI pooling and ROI align

 

Input ：

and RoI Pooling The layer is responsible for collecting proposal, And calculate proposal feature maps, Into the following network . From the picture 2 Can be seen in Rol pooling Layer has a 2 Inputs ：

1. The original feature maps
2. RPN Output proposal boxes（ Different sizes ）

reason ：

 

RoI Pooling principle ：

Let's look at it before we analyze it RoI Pooling Layer Of caffe prototxt The definition of ：

layer {
  name: "roi_pool5"
  type: "ROIPooling"
  bottom: "conv5_3"
  bottom: "rois"
  top: "pool5"
  roi_pooling_param {
    pooled_w: 7
    pooled_h: 7
    spatial_scale: 0.0625 # 1/16
  }
}

There are new parameters pooled_w and pooled_h, Another parameter spatial_scale Readers who read carefully must already know the purpose .RoI Pooling layer forward The process ：

• because proposal It's corresponding to MxN Scale , So first use spatial_scale Parameter maps it back to (M/16)x(N/16) The size of feature map scale ;
• Then each proposal Corresponding feature map The regional level is divided into    The grid of ;
• Every copy of the grid is max pooling Handle .

 

If you take one bin Inside max As this bin The representative value of , Then it is likely to deviate bin Center of , Define precisely the following mask Marginal disadvantage , So the following ROI align Just a change bin Calculation method .

 ROI align