SSD代码解读(一)——Prior Box Layer

SSD系列代码解读:(一) Prior Box

SSD系列代码解读:(二) Data Augmentation

SSD系列代码解读:(三) MultiboxLoss

本部分代码是基于weiliu的ssd,先贴prior_box.hpp代码。

#ifndef CAFFE_PRIORBOX_LAYER_HPP_
#define CAFFE_PRIORBOX_LAYER_HPP_

#include <vector>

#include "caffe/blob.hpp"
#include "caffe/layer.hpp"
#include "caffe/proto/caffe.pb.h"

namespace caffe {

/**
 * @brief Generate the prior boxes of designated sizes and aspect ratios across
 *        all dimensions @f$ (H \times W) @f$.
 *
 * Intended for use with MultiBox detection method to generate prior (template).
 *
 * NOTE: does not implement Backwards operation.
 */
template <typename Dtype>
class PriorBoxLayer : public Layer<Dtype> {
 public:
  /**
   * @param param provides PriorBoxParameter prior_box_param,
   *     with PriorBoxLayer options:
   *   - min_size (\b minimum box size in pixels. can be multiple. required!).
   *   - max_size (\b maximum box size in pixels. can be ignored or same as the
   *   # of min_size.).
   *   - aspect_ratio (\b optional aspect ratios of the boxes. can be multiple).
   *   - flip (\b optional bool, default true).
   *     if set, flip the aspect ratio.
   */
  explicit PriorBoxLayer(const LayerParameter& param)
      : Layer<Dtype>(param) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  virtual inline const char* type() const { return "PriorBox"; }
  virtual inline int ExactBottomBlobs() const { return 2; } //bottom_size = 2
  virtual inline int ExactNumTopBlobs() const { return 1; } //top_size = 1

 protected:
  /**
   * @brief Generates prior boxes for a layer with specified parameters.
   *
   * @param bottom input Blob vector (at least 2)
   *   -# @f$ (N \times C \times H_i \times W_i) @f$
   *      the input layer @f$ x_i @f$
   *   -# @f$ (N \times C \times H_0 \times W_0) @f$
   *      the data layer @f$ x_0 @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times 2 \times K*4) @f$ where @f$ K @f$ is the prior numbers
   *   By default, a box of aspect ratio 1 and min_size and a box of aspect
   *   ratio 1 and sqrt(min_size * max_size) are created.
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  /// @brief Not implemented
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
    return;
  }//Backward_cpu直接return

  vector<float> min_sizes_;
  vector<float> max_sizes_;
  vector<float> aspect_ratios_;
  bool flip_;
  int num_priors_; //表示对特征图的每一个点考虑几个prior box
  bool clip_;
  vector<float> variance_;

  int img_w_;
  int img_h_;
  float step_w_;
  float step_h_;

  float offset_;
};

}  // namespace caffe

#endif  // CAFFE_PRIORBOX_LAYER_HPP_

hpp基本没什么内容,backward直接return.下面贴cpp文件

#include <algorithm>
#include <functional>
#include <utility>
#include <vector>

#include "caffe/layers/prior_box_layer.hpp"

//以fc7为例,参数设置如下
  //prior_box_param {
  //  min_size: 60.0
  //  max_size: 111.0
  //  aspect_ratio: 2
  //  aspect_ratio: 3
  //  flip: true
  //  clip: false
  //  variance: 0.1
  //  variance: 0.1
  //  variance: 0.2
  //  variance: 0.2
  //  step: 16
  //  offset: 0.5
  //}

namespace caffe {

template <typename Dtype>
void PriorBoxLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top) {
  const PriorBoxParameter& prior_box_param =
      this->layer_param_.prior_box_param();
  CHECK_GT(prior_box_param.min_size_size(), 0) << "must provide min_size."; //必须要有min_size
  for (int i = 0; i < prior_box_param.min_size_size(); ++i) {
    min_sizes_.push_back(prior_box_param.min_size(i));
    CHECK_GT(min_sizes_.back(), 0) << "min_size must be positive.";
  }
  aspect_ratios_.clear();
  aspect_ratios_.push_back(1.); //先把1压进aspect_ratios,由min_size * min_size确定的正方形prior box
  flip_ = prior_box_param.flip();
  // 接下来这个for的作用是将prior_box_param.aspect_ratio不重复的压入aspect_ratios.若flip = true,把其倒数也压进去
  for (int i = 0; i < prior_box_param.aspect_ratio_size(); ++i) {
    float ar = prior_box_param.aspect_ratio(i);
    bool already_exist = false;
    for (int j = 0; j < aspect_ratios_.size(); ++j) {
      if (fabs(ar - aspect_ratios_[j]) < 1e-6) {
        already_exist = true;
        break;
      }
    }
    if (!already_exist) {
      aspect_ratios_.push_back(ar);
      if (flip_) { // 
        aspect_ratios_.push_back(1./ar);
      }
    }
  }//aspect_ratios含有1,2,1/2,3,1/3.
  num_priors_ = aspect_ratios_.size() * min_sizes_.size(); // num_priors_=5
  if (prior_box_param.max_size_size() > 0) {
    CHECK_EQ(prior_box_param.min_size_size(), prior_box_param.max_size_size());//max_size个数和min一样
    for (int i = 0; i < prior_box_param.max_size_size(); ++i) {
      max_sizes_.push_back(prior_box_param.max_size(i));
      CHECK_GT(max_sizes_[i], min_sizes_[i])
          << "max_size must be greater than min_size.";
      num_priors_ += 1;//只有一个min和max_size,故num_priors_=6,此处的prior box是边长为sqrt(min*max)的正方形
      }
  }
  clip_ = prior_box_param.clip();
  //variance应该是4个,如果是一个,就把这个值压到variance_里面,其他情况则直接把0.1压到variance_里面
  if (prior_box_param.variance_size() > 1) {
    // Must and only provide 4 variance.
    CHECK_EQ(prior_box_param.variance_size(), 4);
    for (int i = 0; i < prior_box_param.variance_size(); ++i) {
      CHECK_GT(prior_box_param.variance(i), 0);
      variance_.push_back(prior_box_param.variance(i));
    }
  } else if (prior_box_param.variance_size() == 1) {
    CHECK_GT(prior_box_param.variance(0), 0);
    variance_.push_back(prior_box_param.variance(0));
  } else {
    // Set default to 0.1.
    variance_.push_back(0.1);
  }
  // prototxt中一般未给定img_h,img_w和img_size,所以img_h,img_w = 0
  if (prior_box_param.has_img_h() || prior_box_param.has_img_w()) {
    CHECK(!prior_box_param.has_img_size())
        << "Either img_size or img_h/img_w should be specified; not both.";
    img_h_ = prior_box_param.img_h();
    CHECK_GT(img_h_, 0) << "img_h should be larger than 0.";
    img_w_ = prior_box_param.img_w();
    CHECK_GT(img_w_, 0) << "img_w should be larger than 0.";
  } else if (prior_box_param.has_img_size()) {
    const int img_size = prior_box_param.img_size();
    CHECK_GT(img_size, 0) << "img_size should be larger than 0.";
    img_h_ = img_size;
    img_w_ = img_size;
  } else {
    img_h_ = 0;
    img_w_ = 0;
  }
  //step赋值给step_h_和step_w_
  if (prior_box_param.has_step_h() || prior_box_param.has_step_w()) {
    CHECK(!prior_box_param.has_step())
        << "Either step or step_h/step_w should be specified; not both.";
    step_h_ = prior_box_param.step_h();
    CHECK_GT(step_h_, 0.) << "step_h should be larger than 0.";
    step_w_ = prior_box_param.step_w();
    CHECK_GT(step_w_, 0.) << "step_w should be larger than 0.";
  } else if (prior_box_param.has_step()) {
    const float step = prior_box_param.step();
    CHECK_GT(step, 0) << "step should be larger than 0.";
    step_h_ = step;
    step_w_ = step; //step赋值给h_和w_
  } else {
    step_h_ = 0;
    step_w_ = 0;
  }

  offset_ = prior_box_param.offset();
}

template <typename Dtype> //Reshape里的英文注释很重要
void PriorBoxLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top) {
  const int layer_width = bottom[0]->width();
  const int layer_height = bottom[0]->height();
  vector<int> top_shape(3, 1);
  // Since all images in a batch has same height and width, we only need to
  // generate one set of priors which can be shared across all images.
  top_shape[0] = 1; //同一层的priors,feature mp_size,img_size,aspect_ratios都一样,和batch无关,所以为1
  // 2 channels. First channel stores the mean of each prior coordinate.
  // Second channel stores the variance of each prior coordinate.
  top_shape[1] = 2;
  //对于1个prior,不管是prior coordinate还是variance都是4,fc7层有19*19*6个prioi
  top_shape[2] = layer_width * layer_height * num_priors_ * 4;
  CHECK_GT(top_shape[2], 0);
  top[0]->Reshape(top_shape); // blob的reshape方法
}

template <typename Dtype>
void PriorBoxLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
    const vector<Blob<Dtype>*>& top) {
  const int layer_width = bottom[0]->width();
  const int layer_height = bottom[0]->height();
  int img_width, img_height;
  if (img_h_ == 0 || img_w_ == 0) {
    img_width = bottom[1]->width();// 300
    img_height = bottom[1]->height();// 300
  } else {
    img_width = img_w_;
    img_height = img_h_;
  }
  float step_w, step_h;
  if (step_w_ == 0 || step_h_ == 0) {
    step_w = static_cast<float>(img_width) / layer_width;
    step_h = static_cast<float>(img_height) / layer_height;
  } else {
    step_w = step_w_;
    step_h = step_h_;
  }
  Dtype* top_data = top[0]->mutable_cpu_data();
  int dim = layer_height * layer_width * num_priors_ * 4; //19*19*6*4
  int idx = 0;
  
  for (int h = 0; h < layer_height; ++h) {
    for (int w = 0; w < layer_width; ++w) {
      float center_x = (w + offset_) * step_w; 
      float center_y = (h + offset_) * step_h; // feature map上的点对应于原图上的位置,offset_=0.5做到了四舍五入
      float box_width, box_height;
      for (int s = 0; s < min_sizes_.size(); ++s) {
        int min_size_ = min_sizes_[s];
        //min_size*min_size的prioi box(60*60)总会有的,需要注意一下的是idx一直在自增
        // first prior: aspect_ratio = 1, size = min_size
        box_width = box_height = min_size_; 
        // xmin
        top_data[idx++] = (center_x - box_width / 2.) / img_width;
        // ymin
        top_data[idx++] = (center_y - box_height / 2.) / img_height;
        // xmax
        top_data[idx++] = (center_x + box_width / 2.) / img_width;
        // ymax
        top_data[idx++] = (center_y + box_height / 2.) / img_height;//私认为除以300的作用是方便处理数据,数据太大不好处理
		
		//设置了max_size的话,再生成一个sqrt(60*111)为边长的prio box
        if (max_sizes_.size() > 0) {
          CHECK_EQ(min_sizes_.size(), max_sizes_.size());
          int max_size_ = max_sizes_[s];
          // second prior: aspect_ratio = 1, size = sqrt(min_size * max_size)
          box_width = box_height = sqrt(min_size_ * max_size_);
          // xmin
          top_data[idx++] = (center_x - box_width / 2.) / img_width;
          // ymin
          top_data[idx++] = (center_y - box_height / 2.) / img_height;
          // xmax
          top_data[idx++] = (center_x + box_width / 2.) / img_width;
          // ymax
          top_data[idx++] = (center_y + box_height / 2.) / img_height;
        }
		
		// 生成aspect_ratio=2,1/2,3,1/3的四种prior box,宽为60*sqrt(aspect_ratio),高为60/sqrt(aspect_ratio)
        // rest of priors
        for (int r = 0; r < aspect_ratios_.size(); ++r) {
          float ar = aspect_ratios_[r];
          if (fabs(ar - 1.) < 1e-6) {
            continue;
          }
          box_width = min_size_ * sqrt(ar);
          box_height = min_size_ / sqrt(ar);
          // xmin
          top_data[idx++] = (center_x - box_width / 2.) / img_width;
          // ymin
          top_data[idx++] = (center_y - box_height / 2.) / img_height;
          // xmax
          top_data[idx++] = (center_x + box_width / 2.) / img_width;
          // ymax
          top_data[idx++] = (center_y + box_height / 2.) / img_height;
        }
      }
    }
  }//top_data依次存储每个prior box坐标信息,因为/300了,所以都会小于1
  
  //clip让prior box的坐标越左界的置0,越右界置1
  // clip the prior's coordidate such that it is within [0, 1]
  if (clip_) {
    for (int d = 0; d < dim; ++d) {
      top_data[d] = std::min<Dtype>(std::max<Dtype>(top_data[d], 0.), 1.);
    }
  }
  //至此,top_data第一个channel的dim个坐标信息已写好,接下来写第二个channel的dim个variance_信息
  
  // set the variance.
  top_data += top[0]->offset(0, 1);//调用了blob的offset函数,计算出第二个channel初始位置的偏移量,使后面使用top_data[0]就可以设置variance_
  if (variance_.size() == 1) {
    caffe_set<Dtype>(dim, Dtype(variance_[0]), top_data);
  } else {
    int count = 0;
    for (int h = 0; h < layer_height; ++h) {
      for (int w = 0; w < layer_width; ++w) {
        for (int i = 0; i < num_priors_; ++i) {
          for (int j = 0; j < 4; ++j) {
            top_data[count] = variance_[j];
            ++count;
          }
        }
      }
    }
  }
}

INSTANTIATE_CLASS(PriorBoxLayer);
REGISTER_LAYER_CLASS(PriorBox);

}  // namespace caffe

盗用知乎上的一张图来帮助理解:先将feature map上的每个点对应到300*300的img上,作为中心点.依此中心点做出num_prior个prior box,再把各个超出边界的box拉回来(前提是clip=true).

注释的比较多,在此总结一下:

1.num_prior=aspect_ratio个数*2 (flip=true)+ min_size为边长的1个正方形+sqrt(min*max_size)为边长的1个正方形;

2.关于img_h,img_w,step_w等等,参考caffe.proto文件.

// By default, we calculate img_height, img_width, step_x, step_y based on
  // bottom[0] (feat) and bottom[1] (img). Unless these values are explicitely
  // provided.
  // Explicitly provide the img_size.
  optional uint32 img_size = 7;
  // Either img_size or img_h/img_w should be specified; not both.
  optional uint32 img_h = 8;
  optional uint32 img_w = 9;

  // Explicitly provide the step size.
  optional float step = 10;
  // Either step or step_h/step_w should be specified; not both.
  optional float step_h = 11;
  optional float step_w = 12;

这些参数是可给可不给的,网络可以自己获得,关于step再说一句:到该层下采样了n次,step=2^n(stride=2),和faster rcnn的含义一样.

3.关于varience,参见github. 除以varience是对预测框和真实框的误差进行放大,从而增大loss,增大梯度,加快收敛.

4.整个prior层以feature map和data层作为输入,为feature map每个点考虑num_prior个prioi box,输出shape为(1,2,layer_height * layer_width * num_priors_ * 4),也就是2个channel,第一个channel存放每个prioi box映射回原图的位置信息,第二个channel存放每个prioi box的varience信息.其实prior box就和anchor差不多,只不过前者在多scale的featurp map上获得且个数不为9.

5.关于min_size的设计问题,论文中给出了公式:

ssd的m=6(取了6个层做检测),第一层一般单独给出,第二层feature map对应的min_size=S1*300, max_size=S2*300;第三层min_size=S2*300, max_size=S3*300;以此类推...

代码中的实现略有不同, ssd_voc_pascal.py中分母除的是m-2,第一层的scale分别是0.1和0.2;ssd_coco中smin=0.15.max=0.9,且分母除的也是m-2.可以看出针对coco中小物体增多的情况下,采用了更小的smin,这也是ssd应用在自己的data domain中需要考虑的问题.

# ssd_pascal.py
min_ratio = 20
max_ratio = 90
step = int(math.floor((max_ratio - min_ratio) / (len(mbox_source_layers) - 2)))
min_sizes = []
max_sizes = []
for ratio in xrange(min_ratio, max_ratio + 1, step):
  min_sizes.append(min_dim * ratio / 100.)
  max_sizes.append(min_dim * (ratio + step) / 100.)
min_sizes = [min_dim * 10 / 100.] + min_sizes
max_sizes = [min_dim * 20 / 100.] + max_sizes

 

最后,越看SSD,越觉得RGB真的牛, RPN网络是真的里程碑式设计.

 

展开阅读全文

没有更多推荐了,返回首页