CPM中data_transformer.cpp代码片段学习

程序功能

caffe，CPM程序中，进行数据预处理部分，预处理的方式:
(1)图像对比度增强clahe
(2)flip
(3)rote
(4)中间点
(5)crop
(6)color Jittering:颜色增强（颜色增强、饱和度、对比度）

#ifdef USE_OPENCV
#include <opencv2/core/core.hpp>
//#include <opencv2/opencv.hpp>
#include <opencv2/contrib/contrib.hpp>
#include <opencv2/highgui/highgui.hpp>
#endif  // USE_OPENCV

#include <iostream>
#include <algorithm>
#include <fstream>
using namespace cv;
using namespace std;

#include <string>
#include <sstream>
#include <vector>

#include "caffe/data_transformer.hpp"
#include "caffe/util/io.hpp"
#include "caffe/util/math_functions.hpp"
#include "caffe/util/rng.hpp"

namespace caffe {
   
/*参数：data，idx，pf,len
  功能：
*/
template<typename Dtype>
void DecodeFloats(const string& data, size_t idx, Dtype* pf, size_t len) {
   
  memcpy(pf, const_cast<char*>(&data[idx]), len * sizeof(Dtype));
}
/*参数：引用变量（？）data，(size_t用来记录大小的数据类型)idx
  功能：连接data[idx]，data[idx+1]，data[idx+2]，结果为一长串字符串，保存着dataset name
*/
string DecodeString(const string& data, size_t idx) {
   
  string result = "";
  int i = 0;
  while(data[idx+i] != 0){
   
    result.push_back(char(data[idx+i]));
    i++;
  }
  return result;
}
/*参数：MetaData,data,offset3(=3*row*col),offset1(data_width)
  功能：读取dataset name ； height, width; img_size；isValidation；
       numOtherPeople； people_index；annolist_index；
       write_number;
       total_write_number;以及坐标
       记录训练次数
*/
template<typename Dtype>
void DataTransformer<Dtype>::ReadMetaData(MetaData& meta, const string& data, size_t offset3, size_t offset1) {
    //very specific to genLMDB.py
  //ReadMetaData（）是类DataTransformer的成员函数，通过::实现ReadMetaData（）函数功能
  //Metadata是容器
  // ------------------- Dataset name ----------------------
  meta.dataset = DecodeString(data, offset3);
  // ------------------- Image Dimension -------------------
  float height, width;
  DecodeFloats(data, offset3+offset1, &height, 1);//参数(data, idx, Dtype* pf, len=1)
  DecodeFloats(data, offset3+offset1+4, &width, 1); //函数memcpy(pf, const_cast<char*>(&data[offset3+offset1]), sizeof(Dtype));
  //pf（&height）<- &data[offset3+offset1]
  meta.img_size = Size(width, height);
  // ----------- Validation, nop, counters -----------------
  meta.isValidation = (data[offset3+2*offset1]==0 ? false : true);//是否为验证集（data[]里面存储了验证集标签信息，从offset3+2*offset1开始）
  meta.numOtherPeople = (int)data[offset3+2*offset1+1];//图片中其他人的数量
  meta.people_index = (int)data[offset3+2*offset1+2];每个人的索引
  float annolist_index;
  DecodeFloats(data, offset3+2*offset1+3, &annolist_index, 1);
  meta.annolist_index = (int)annolist_index;//【？】&annolist_index<- &data[offset3+2*offset1+3]
  float write_number;
  DecodeFloats(data, offset3+2*offset1+7, &write_number, 1);
  meta.write_number = (int)write_number;//【当前所写入的第几张图片】&write_number<- &data[offset3+2*offset1+7]
  float total_write_number;
  DecodeFloats(data, offset3+2*offset1+11, &total_write_number, 1);
  meta.total_write_number = (int)total_write_number;//【样本总数】

  // count epochs according to counters
  //在神经网络中数据集会被训练多次，在多次训练时会需要记录第几次重复做【预处理-->输入网络】这个过程，即训练次数；
  //每次训练时，都会经过这个预处理，每次处理的结果不一样
  static int cur_epoch = -1;
  if(meta.write_number == 0){
   
    cur_epoch++;
  }
  meta.epoch = cur_epoch;
  if(meta.write_number % 1000 == 0){
   
    LOG(INFO) << "dataset: " << meta.dataset <<"; img_size: " << meta.img_size
        << "; meta.annolist_index: " << meta.annolist_index << "; meta.write_number: " << meta.write_number
        << "; meta.total_write_number: " << meta.total_write_number << "; meta.epoch: " << meta.epoch;
  }//每读1000张图片输出以上信息
  if(param_.aug_way() == "table" && !is_table_set){
   
    SetAugTable(meta.total_write_number);
    is_table_set = true;
  }//用于列表每一张图的增强处理次数

  // ------------------- objpos 人物中心点-----------------------
  DecodeFloats(data, offset3+3*offset1, &meta.objpos.x, 1);//&meta.objpos.x<- &data[offset3+3*offset1]
  DecodeFloats(data, offset3+3*offset1+4, &meta.objpos.y, 1);
  meta.objpos -= Point2f(1,1);//由于meta里面的数据是以matlab为标准（从1开头），所以在c++中坐标应该减一
  // ------------ scale_self, joint_self --------------
  DecodeFloats(data, offset3+4*offset1, &meta.scale_self, 1);
  meta.joint_self.joints.resize(np_in_lmdb);//关节resize，调整容器的大小，使其包含”关节个数“个元素。
  meta.joint_self.isVisible.resize(np_in_lmdb);
  for(int i=0; i<np_in_lmdb; i++){
   //关节坐标的赋值与转换，标签是否可见、超出坐标范围
    DecodeFloats(data, offset3+5*offset1+4*i, &meta.joint_self.joints[i].x, 1);
    DecodeFloats(data, offset3+6*offset1+4*i, &meta.joint_self.joints[i].y, 1);//每个人关节的横纵坐标
    meta.joint_self.joints[i] -= Point2f(1,1); //from matlab 1-index to c++ 0-index
    float isVisible;
    DecodeFloats(data, offset3+7*offset1+4*i, &isVisible, 1);
    meta.joint_self.isVisible[i] = (isVisible == 0) ? 0 : 1;//可见为0，不可见为1
    if(meta.joint_self.joints[i].x < 0 || meta.joint_self.joints[i].y < 0 ||
       meta.joint_self.joints[i].x >= meta.img_size.width || meta.joint_self.joints[i].y >= meta.img_size.height){
   
      meta.joint_self.isVisible[i] = 2; // 2 means cropped, 0 means occluded by still on image
    }//1:不可见
    // 2：要裁剪
    // 0：可见且不须裁剪
    //LOG(INFO) << meta.joint_self.joints[i].x << " " << meta.joint_self.joints[i].y << " " << meta.joint_self.isVisible[i];
  }
  
  //others (7 lines loaded)
  meta.objpos_other.resize(meta.numOtherPeople);//其他人的中心点，高度，关节的容器结构大小调整
  meta.scale_other.resize(meta.numOtherPeople);
  meta.joint_others.resize(meta.numOtherPeople);
  for(int p=0; p<meta.numOtherPeople; p++){
   //（0~其他人数量）
    DecodeFloats(data, offset3+(8+p)*offset1, &meta.objpos_other[p].x, 1);//中心点坐标赋值
    DecodeFloats(data, offset3+(8+p)*offset1+4, &meta.objpos_other[p].y, 1);
    meta.objpos_other[p] -= Point2f(1,1);//坐标转换
    DecodeFloats(data, offset3+(8+meta.numOtherPeople)*offset1+4*p, &meta.scale_other[p], 1);
  }
  //8 + numOtherPeople lines loaded
  for(int p=0; p<meta.numOtherPeople; p++){
   //遍历每个”其它“人
    meta.joint_others[p].joints.resize(np_in_lmdb);//其他人关节点的可见性，joints，调整容器大小
    meta.joint_others[p].isVisible.resize(np_in_lmdb);
    for(int i=0; i<np_in_lmdb; i++){
   
      DecodeFloats(data, offset3+(9+meta.numOtherPeople+3*p)*offset1+4*i, &meta.joint_others[p].joints[i].x, 1);
      DecodeFloats(data, offset3+(9+meta.numOtherPeople+3*p+1)*offset1+4*i, &meta.joint_others[p].joints[i].y, 1);
      meta.joint_others[p].joints[i] -= Point2f(1,1);
      float isVisible;
      DecodeFloats(data, offset3+(9+meta.numOtherPeople+3*p+2)*offset1+4*i, &isVisible, 1);
      meta.joint_others[p].isVisible[i] = (isVisible == 0) ? 0 : 1;
      if(meta.joint_others[p].joints[i].x < 0 || meta.joint_others[p].joints[i].y < 0 ||
         meta.joint_others[p].joints[i].x >= meta.img_size.width || meta.joint_others[p].joints[i].y >= meta.img_size.height){
   
        meta.joint_others[p].isVisible[i] = 2; // 2 means cropped, 1 means occluded by stil