faster-rcnn中ROI_POOIING層的解讀

在沒有出現sppnet以前，RCNN使用corp和warp來對圖片進行大小調整，這種操做會形成圖片信息失真和信息丟失。sppnet這個模型推出來以後（關於這個網絡的描述，能夠看看以前寫的一篇理解：http://www.cnblogs.com/gongxijun/p/7172134.html），rg大神沿用了sppnet的思路到他的下一個模型中fast-rcnn中，可是roi_pooling和sppnet的思路雖然相同，可是實現方式仍是不一樣的.咱們看一下網絡參數:

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
}

結合源代碼，做者藉助了sppnet的空域金字塔pool方式，可是和sppnet並不一樣的是，做者在這裏只使用了(pooled_w,pooled_h)這個尺度，來將獲得的每個特徵圖分紅(pooled_w,pooled_h)，而後對每一塊進行max_pooling取值，最後獲得一個n*7*7固定大小的特徵圖。

  1 // ------------------------------------------------------------------
  2 // Fast R-CNN
  3 // Copyright (c) 2015 Microsoft
  4 // Licensed under The MIT License [see fast-rcnn/LICENSE for details]
  5 // Written by Ross Girshick
  6 // ------------------------------------------------------------------
  7 
  8 #include <cfloat>
  9 
 10 #include "caffe/fast_rcnn_layers.hpp"
 11 
 12 using std::max;
 13 using std::min;
 14 using std::floor;
 15 using std::ceil;
 16 
 17 namespace caffe {
 18 
 19 template <typename Dtype>
 20 void ROIPoolingLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
 21       const vector<Blob<Dtype>*>& top) {
 22   ROIPoolingParameter roi_pool_param = this->layer_param_.roi_pooling_param();
 23   CHECK_GT(roi_pool_param.pooled_h(), 0)
 24       << "pooled_h must be > 0";
 25   CHECK_GT(roi_pool_param.pooled_w(), 0)
 26       << "pooled_w must be > 0";
 27   pooled_height_ = roi_pool_param.pooled_h(); //定義網絡的大小
 28   pooled_width_ = roi_pool_param.pooled_w();
 29   spatial_scale_ = roi_pool_param.spatial_scale();
 30   LOG(INFO) << "Spatial scale: " << spatial_scale_;
 31 }
 32 
 33 template <typename Dtype>
 34 void ROIPoolingLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
 35       const vector<Blob<Dtype>*>& top) {
 36   channels_ = bottom[0]->channels();
 37   height_ = bottom[0]->height();
 38   width_ = bottom[0]->width();
 39   top[0]->Reshape(bottom[1]->num(), channels_, pooled_height_,
 40       pooled_width_);
 41   max_idx_.Reshape(bottom[1]->num(), channels_, pooled_height_,
 42       pooled_width_);
 43 }
 44 
 45 template <typename Dtype>
 46 void ROIPoolingLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
 47       const vector<Blob<Dtype>*>& top) {
 48   const Dtype* bottom_data = bottom[0]->cpu_data();
 49   const Dtype* bottom_rois = bottom[1]->cpu_data();//獲取roidb信息(n,x1,y1,x2,y2)
 50   // Number of ROIs
 51   int num_rois = bottom[1]->num();//候選目標的個數
 52   int batch_size = bottom[0]->num();//特徵圖的維度,vgg16的conv5以後爲512
 53   int top_count = top[0]->count();//須要輸出的值個數
 54   Dtype* top_data = top[0]->mutable_cpu_data();
 55   caffe_set(top_count, Dtype(-FLT_MAX), top_data);
 56   int* argmax_data = max_idx_.mutable_cpu_data();
 57   caffe_set(top_count, -1, argmax_data);
 58 
 59   // For each ROI R = [batch_index x1 y1 x2 y2]: max pool over R
 60   for (int n = 0; n < num_rois; ++n) {
 61     int roi_batch_ind = bottom_rois[0];
 62     int roi_start_w = round(bottom_rois[1] * spatial_scale_);//縮小１６倍，將候選區域在原始座標中的位置，映射到conv_5特徵圖上
 63     int roi_start_h = round(bottom_rois[2] * spatial_scale_);
 64     int roi_end_w = round(bottom_rois[3] * spatial_scale_);
 65     int roi_end_h = round(bottom_rois[4] * spatial_scale_);
 66     CHECK_GE(roi_batch_ind, 0);
 67     CHECK_LT(roi_batch_ind, batch_size);
 68 
 69     int roi_height = max(roi_end_h - roi_start_h + 1, 1);//獲得候選區域在特徵圖上的大小
 70     int roi_width = max(roi_end_w - roi_start_w + 1, 1);
 71     const Dtype bin_size_h = static_cast<Dtype>(roi_height)
 72                              / static_cast<Dtype>(pooled_height_);//計算若是須要劃分紅(pooled_height_，pooled_weight_)這麼多塊，那麼每個塊的大小(bin_size_w,bin_size_h);
 73     const Dtype bin_size_w = static_cast<Dtype>(roi_width)
 74                              / static_cast<Dtype>(pooled_width_);
 75 
 76     const Dtype* batch_data = bottom_data + bottom[0]->offset(roi_batch_ind);//獲取當前維度的特徵圖數據，好比一共有(n,x1,x2,x3,x4)的數據，拿到第一塊特徵圖的數據
 77 
 78     for (int c = 0; c < channels_; ++c) {
 79       for (int ph = 0; ph < pooled_height_; ++ph) {
 80         for (int pw = 0; pw < pooled_width_; ++pw) {
 81           // Compute pooling region for this output unit:
 82           //  start (included) = floor(ph * roi_height / pooled_height_)
 83           //  end (excluded) = ceil((ph + 1) * roi_height / pooled_height_)
 84           int hstart = static_cast<int>(floor(static_cast<Dtype>(ph)
 85                                               * bin_size_h)); //計算每一塊的位置
 86           int wstart = static_cast<int>(floor(static_cast<Dtype>(pw)
 87                                               * bin_size_w));
 88           int hend = static_cast<int>(ceil(static_cast<Dtype>(ph + 1)
 89                                            * bin_size_h));
 90           int wend = static_cast<int>(ceil(static_cast<Dtype>(pw + 1)
 91                                            * bin_size_w));
 92 
 93           hstart = min(max(hstart + roi_start_h, 0), height_);
 94           hend = min(max(hend + roi_start_h, 0), height_);
 95           wstart = min(max(wstart + roi_start_w, 0), width_);
 96           wend = min(max(wend + roi_start_w, 0), width_);
 97 
 98           bool is_empty = (hend <= hstart) || (wend <= wstart);
 99 
100           const int pool_index = ph * pooled_width_ + pw;
101           if (is_empty) {
102             top_data[pool_index] = 0;
103             argmax_data[pool_index] = -1;
104           }
105 
106           for (int h = hstart; h < hend; ++h) {
107             for (int w = wstart; w < wend; ++w) {
108               const int index = h * width_ + w;
109               if (batch_data[index] > top_data[pool_index]) {
110                 top_data[pool_index] = batch_data[index];　//在取每一塊中的最大值，就是max_pooling操做.
111                 argmax_data[pool_index] = index;
112               }
113             }
114           }
115         }
116       }
117       // Increment all data pointers by one channel
118       batch_data += bottom[0]->offset(0, 1);
119       top_data += top[0]->offset(0, 1);
120       argmax_data += max_idx_.offset(0, 1);
121     }
122     // Increment ROI data pointer
123     bottom_rois += bottom[1]->offset(1);
124   }
125 }
126 
127 template <typename Dtype>
128 void ROIPoolingLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
129       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
130   NOT_IMPLEMENTED;
131 }
132 
133 
134 #ifdef CPU_ONLY
135 STUB_GPU(ROIPoolingLayer);
136 #endif
137 
138 INSTANTIATE_CLASS(ROIPoolingLayer);
139 REGISTER_LAYER_CLASS(ROIPooling);
140 
141 }  // namespace caffe

進過以上的操做事後，就獲得了固定大小的特徵圖啦，而後就能夠進行全鏈接操做了. 希望我說明白了.

 

－－－完.