ssd网络详解之detection output layer

ssd网络详解之detection output layer

本文原创，转载请引用https://blog.csdn.net/dan_teng/article/details/81561783

闲话少说，本文基本结构：首先介绍detection output 这一层的基本理解，之后给出ssd所有代码的详细注释，最后给出caffe中该层各个参数的定义和默认值。

detection out layer是ssd网络最后一层，用于整合预选框、预选框偏移以及得分三项结果，最终输出满足条件的目标检测框、目标的label和得分。
输入方面，mbox_priorbox是网络各个priorbox层输出concat后的结果（priorbox解析点这里），相当于把所有预选框放到一起；mbox_loc是在预选框的基础上的偏移量；mbox_conf_flatten就是每个类别在各个框上的得分。
输出大小为[1, 1, x, 7]，其中x是最后保留的框的个数，最后一维存放的数据为：
[image_id, label, confidence, xmin, ymin, xmax, ymax]

计算思路：
1）对bottom层的location、confidence和priorbox进行解析，放到vector中
2）对每个priorbox进行解码。所谓解码其实就是整合输入层。前面说到过了，输出需要给出每个目标的检测框，但是输入是预选框和偏移量，这里要做的就是计算出最终的检测框。解码需要考虑priorbox编码方式，共三种情况。

假设检测框用b表示(存储内容：b_xmin, b_ymin, b_xmax, b_ymax)，预选框用p表示(存储内容：p_xmin, p_ymin, p_xmax, p_ymax)，偏移量用t表示(存储内容：t_x, t_y, t_height, t_width)。
b和p的宽高分别用x和y的最大最小值减一下得到，中心点的值用最大最小值相加除以2得到。
那么在每种类型中，编码公式分别为：

• CodeType_CORNER：

                                t = b - p（每个维度一样）
  

• CodeType_CENTER_SIZE：

                               t_x = (b_center_x – p_center_x) / p_width （t_y同理）
                               t_height = log(b_height / p_height) （t_width同理）
  

• CodeType_CORNER_SIZE：

                                t_x = (b_x – p_x) / p_width
                                t_y = (b_y – p_y) / p_height
  

解码时求取b的各个值就可以。如果需要添加variance的值，将t与variance相乘即可。
以center_size编码方式为例：

               b_center_x = t_x * p_width + p_center_x
               b_center_y = t_y * p_height + p_center_y
               b_width = exp(t_x) * p_width
               b_height = exp(t_y) * p_height

如果需要添加variance：

               b_center_x = t_x *prior_variance[0]* p_width + p_center_x
               b_center_y = t_y *prior_variance[1] * p_height + p_center_y
               b_width = exp(prior_variance[2] * t_x) * p_width
               b_height = exp(prior_variance[3] * t_y) * p_height

据此分别计算出b_xmin, b_ymin, b_xmax, b_ymax即可。详细可参见代码

3） Non-Maximum Suppression非极大值抑制
检测算法给出的box往往有很多，如下图所示，多个检测框其实框出的是一个目标，nms就是一个目标保留一个最优框。抑制的过程是一个迭代-遍历-消除的过程。

（图片来源：https://blog.csdn.net/shuzfan/article/details/52711706）

给定处理前的集合：预选结合，处理后的集合keep集合

首先，将预选集合所有框按照得分高低进行排序，选中得分最高的框，从预选集合移出放到keep集合中；

接下来进行迭代：
*从当前预选集合移出得分最高的框，用它与keep集合每个框计算交并比：
*超过阈值说明二者重复很多，框住的应该是同一个东西，不放到keep集合中；
*如果与keep集合中每个框交并比都小于阈值，说明当前框框住的是一个新目标，应该放到keep中。

迭代下去，直到预选集合为空，那么keep集合中留下的就是检出的所有目标的检测框。

jaccard overlap
这里补充介绍一下ssd网络中的jaccard overlap。
jaccard overlap其实就是交并比，简单说起来就是两个检测框重合的面积（相交的部分）除以两个检测框并在一起的面积（面积之和减去重合部分），用公式表示为

$$J(A,B) = \frac{|A\cap B|}{|A\cup B|}$$
J为0说明两个框一点没有重合，为1说明完全重合

4）按照输出大小要求输出结果

代码详解：

注意：这里给出了detection_output_layer.cpp中的代码，但是代码里用到了一些函数，这些函数放在了
ssd/src/caffe/util/bbox_util.cpp中

#include <algorithm>
#include <fstream>  // NOLINT(readability/streams)
#include <map>
#include <string>
#include <utility>
#include <vector>

#include "boost/filesystem.hpp"
#include "boost/foreach.hpp"

#include "caffe/layers/detection_output_layer.hpp"

namespace caffe {

template <typename Dtype>
void DetectionOutputLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top) {
  const DetectionOutputParameter& detection_output_param =
      this->layer_param_.detection_output_param();
  CHECK(detection_output_param.has_num_classes()) << "Must specify num_classes";
  num_classes_ = detection_output_param.num_classes();// 类别数量
  share_location_ = detection_output_param.share_location();
  num_loc_classes_ = share_location_ ? 1 : num_classes_;
  background_label_id_ = detection_output_param.background_label_id();
  code_type_ = detection_output_param.code_type();// 编码类型
  variance_encoded_in_target_ =
      detection_output_param.variance_encoded_in_target();
  keep_top_k_ = detection_output_param.keep_top_k(); // 保留框最大数量
  confidence_threshold_ = detection_output_param.has_confidence_threshold() ?
      detection_output_param.confidence_threshold() : -FLT_MAX;// 得分阈值
  // Parameters used in nms.
  nms_threshold_ = detection_output_param.nms_param().nms_threshold();
  CHECK_GE(nms_threshold_, 0.) << "nms_threshold must be non negative.";
  eta_ = detection_output_param.nms_param().eta();
  CHECK_GT(eta_, 0.);
  CHECK_LE(eta_, 1.);
  top_k_ = -1;
  if (detection_output_param.nms_param().has_top_k()) {
    top_k_ = detection_output_param.nms_param().top_k();
  }
  const SaveOutputParameter& save_output_param =
      detection_output_param.save_output_param();
  output_directory_ = save_output_param.output_directory();
  if (!output_directory_.empty()) {
    if (boost::filesystem::is_directory(output_directory_)) {
      boost::filesystem::remove_all(output_directory_);
    }
    if (!boost::filesystem::create_directories(output_directory_)) {
        LOG(WARNING) << "Failed to create directory: " << output_directory_;
    }
  }
  output_name_prefix_ = save_output_param.output_name_prefix();
  need_save_ = output_directory_ == "" ? false : true;
  output_format_ = save_output_param.output_format();
  if (save_output_param.has_label_map_file()) {
    string label_map_file = save_output_param.label_map_file();
    if (label_map_file.empty()) {
      // Ignore saving if there is no label_map_file provided.
      LOG(WARNING) << "Provide label_map_file if output results to files.";
      need_save_ = false;
    } else {
      LabelMap label_map;
      CHECK(ReadProtoFromTextFile(label_map_file, &label_map))
          << "Failed to read label map file: " << label_map_file;
      CHECK(MapLabelToName(label_map, true, &label_to_name_))
          << "Failed to convert label to name.";
      CHECK(MapLabelToDisplayName(label_map, true, &label_to_display_name_))
          << "Failed to convert label to display name.";
    }
  } else {
    need_save_ = false;
  }
  if (save_output_param.has_name_size_file()) {
    string name_size_file = save_output_param.name_size_file();
    if (name_size_file.empty()) {
      // Ignore saving if there is no name_size_file provided.
      LOG(WARNING) << "Provide name_size_file if output results to files.";
      need_save_ = false;
    } else {
      std::ifstream infile(name_size_file.c_str());
      CHECK(infile.good())
          << "Failed to open name size file: " << name_size_file;
      // The file is in the following format:
      //    name height width
      //    ...
      string name;
      int height, width;
      while (infile >> name >> height >> width) {
        names_.push_back(name);
        sizes_.push_back(std::make_pair(height, width));
      }
      infile.close();
      if (save_output_param.has_num_test_image()) {
        num_test_image_ = save_output_param.num_test_image();
      } else {
        num_test_image_ = names_.size();
      }
      CHECK_LE(num_test_image_, names_.size());
    }
  } else {
    need_save_ = false;
  }
  has_resize_ = save_output_param.has_resize_param();
  if (has_resize_) {
    resize_param_ = save_output_param.resize_param();
  }
  name_count_ = 0;
  visualize_ = detection_output_param.visualize();
  if (visualize_) {
    visualize_threshold_ = 0.6;
    if (detection_output_param.has_visualize_threshold()) {
      visualize_threshold_ = detection_output_param.visualize_threshold();
    }
    data_transformer_.reset(
        new DataTransformer<Dtype>(this->layer_param_.transform_param(),
                                   this->phase_));
    data_transformer_->InitRand();
    save_file_ = detection_output_param.save_file();
  }
  bbox_preds_.ReshapeLike(*(bottom[0]));
  if (!share_location_) {
    bbox_permute_.ReshapeLike(*(bottom[0]));
  }
  conf_permute_.ReshapeLike(*(bottom[1]));
}
// 输出大小为[1, 1, x, 7]
// 最后一维7指的是：[image_id, label, confidence, xmin, ymin, xmax, ymax]
template <typename Dtype>
void DetectionOutputLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top) {
  if (need_save_) {
    CHECK_LE(name_count_, names_.size());
    if (name_count_ % num_test_image_ == 0) {
      // Clean all outputs.
      if (output_format_ == "VOC") {
        boost::filesystem::path output_directory(output_directory_);
        for (map<int, string>::iterator it = label_to_name_.begin();
             it != label_to_name_.end(); ++it) {
          if (it->first == background_label_id_) {
            continue;
          }
          std::ofstream outfile;
          boost::filesystem::path file(
              output_name_prefix_ + it->second + ".txt");
          boost::filesystem::path out_file = output_directory / file;
          outfile.open(out_file.string().c_str(), std::ofstream::out);
        }
      }
    }
  }
  CHECK_EQ(bottom[0]->num(), bottom[1]->num());
  if (bbox_preds_.num() != bottom[0]->num() ||
      bbox_preds_.count(1) != bottom[0]->count(1)) {
    bbox_preds_.ReshapeLike(*(bottom[0]));
  }
  if (!share_location_ && (bbox_permute_.num() != bottom[0]->num() ||
      bbox_permute_.count(1) != bottom[0]->count(1))) {
    bbox_permute_.ReshapeLike(*(bottom[0]));
  }
  if (conf_permute_.num() != bottom[1]->num() ||
      conf_permute_.count(1) != bottom[1]->count(1)) {
    conf_permute_.ReshapeLike(*(bottom[1]));
  }
  num_priors_ = bottom[2]->height() / 4;
  CHECK_EQ(num_priors_ * num_loc_classes_ * 4, bottom[0]->channels())
      << "Number of priors must match number of location predictions.";
  CHECK_EQ(num_priors_ * num_classes_, bottom[1]->channels())
      << "Number of priors must match number of confidence predictions.";
  // num() and channels() are 1.
  vector<int> top_shape(2, 1);
  // Since the number of bboxes to be kept is unknown before nms, we manually
  // set it to (fake) 1.
  top_shape.push_back(1);
  // Each row is a 7 dimension vector, which stores
  // [image_id, label, confidence, xmin, ymin, xmax, ymax]
  top_shape.push_back(7);
  top[0]->Reshape(top_shape);
}

template <typename Dtype>
void DetectionOutputLayer<Dtype>::Forward_cpu(
    const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
  const Dtype* loc_data = bottom[0]->cpu_data();
  const Dtype* conf_data = bottom[1]->cpu_data();
  const Dtype* prior_data = bottom[2]->cpu_data();
  const int num = bottom[0]->num();

  // Retrieve all location predictions.
  vector<LabelBBox> all_loc_preds;
  // 处理偏移量数据
  GetLocPredictions(loc_data, num, num_priors_, num_loc_classes_,
                    share_location_, &all_loc_preds);

  // Retrieve all confidences.
  vector<map<int, vector<float> > > all_conf_scores;
  // 处理得分数据
  GetConfidenceScores(conf_data, num, num_priors_, num_classes_,
                      &all_conf_scores);

  // Retrieve all prior bboxes. It is same within a batch since we assume all
  // images in a batch are of same dimension.
  vector<NormalizedBBox> prior_bboxes;
  vector<vector<float> > prior_variances;
  // 处理预选框数据
  GetPriorBBoxes(prior_data, num_priors_, &prior_bboxes, &prior_variances);

  // Decode all loc predictions to bboxes.
  vector<LabelBBox> all_decode_bboxes;
  const bool clip_bbox = false;
  // 解码
  DecodeBBoxesAll(all_loc_preds, prior_bboxes, prior_variances, num,
                  share_location_, num_loc_classes_, background_label_id_,
                  code_type_, variance_encoded_in_target_, clip_bbox,
                  &all_decode_bboxes);

  int num_kept = 0;
  vector<map<int, vector<int> > > all_indices;
  for (int i = 0; i < num; ++i) {
    const LabelBBox& decode_bboxes = all_decode_bboxes[i];
    const map<int, vector<float> >& conf_scores = all_conf_scores[i];
    map<int, vector<int> > indices;
    int num_det = 0;
    for (int c = 0; c < num_classes_; ++c) {
      if (c == background_label_id_) {
        // Ignore background class.
        continue;
      }
      if (conf_scores.find(c) == conf_scores.end()) {
        // Something bad happened if there are no predictions for current label.
        LOG(FATAL) << "Could not find confidence predictions for label " << c;
      }
      const vector<float>& scores = conf_scores.find(c)->second;
      int label = share_location_ ? -1 : c;
      if (decode_bboxes.find(label) == decode_bboxes.end()) {
        // Something bad happened if there are no predictions for current label.
        LOG(FATAL) << "Could not find location predictions for label " << label;
        continue;
      }
      const vector<NormalizedBBox>& bboxes = decode_bboxes.find(label)->second;
      // 非极大值抑制
      ApplyNMSFast(bboxes, scores, confidence_threshold_, nms_threshold_, eta_,
          top_k_, &(indices[c]));
      num_det += indices[c].size();
    }
    // 处理后有效数据量大于最后输出量，那就取得分最高的前keep_top_k个检测框
    if (keep_top_k_ > -1 && num_det > keep_top_k_) {
      vector<pair<float, pair<int, int> > > score_index_pairs;
      for (map<int, vector<int> >::iterator it = indices.begin();
           it != indices.end(); ++it) {
        int label = it->first;
        const vector<int>& label_indices = it->second;
        if (conf_scores.find(label) == conf_scores.end()) {
          // Something bad happened for current label.
          LOG(FATAL) << "Could not find location predictions for " << label;
          continue;
        }
        const vector<float>& scores = conf_scores.find(label)->second;
        for (int j = 0; j < label_indices.size(); ++j) {
          int idx = label_indices[j];
          CHECK_LT(idx, scores.size());
          score_index_pairs.push_back(std::make_pair(
                  scores[idx], std::make_pair(label, idx)));
        }
      }
      // Keep top k results per image.
      std::sort(score_index_pairs.begin(), score_index_pairs.end(),
                SortScorePairDescend<pair<int, int> >);
      score_index_pairs.resize(keep_top_k_);
      // Store the new indices.
      map<int, vector<int> > new_indices;
      for (int j = 0; j < score_index_pairs.size(); ++j) {
        int label = score_index_pairs[j].second.first;
        int idx = score_index_pairs[j].second.second;
        new_indices[label].push_back(idx);
      }
      all_indices.push_back(new_indices);
      num_kept += keep_top_k_;
    } else {
      all_indices.push_back(indices);
      num_kept += num_det;
    }
  }

  vector<int> top_shape(2, 1);
  top_shape.push_back(num_kept);
  top_shape.push_back(7);
  Dtype* top_data;
  // 没有检测到目标
  if (num_kept == 0) {
    LOG(INFO) << "Couldn't find any detections";
    top_shape[2] = num;
    top[0]->Reshape(top_shape);
    top_data = top[0]->mutable_cpu_data();
    caffe_set<Dtype>(top[0]->count(), -1, top_data);
    // Generate fake results per image.
    for (int i = 0; i < num; ++i) {
      top_data[0] = i;
      top_data += 7;
    }
  } else {// 检测到目标
    top[0]->Reshape(top_shape);
    top_data = top[0]->mutable_cpu_data();
  }
  // 检测到目标的处理
  int count = 0;
  boost::filesystem::path output_directory(output_directory_);
  for (int i = 0; i < num; ++i) {
    const map<int, vector<float> >& conf_scores = all_conf_scores[i];
    const LabelBBox& decode_bboxes = all_decode_bboxes[i];
    for (map<int, vector<int> >::iterator it = all_indices[i].begin();
         it != all_indices[i].end(); ++it) {
      int label = it->first;
      if (conf_scores.find(label) == conf_scores.end()) {
        // Something bad happened if there are no predictions for current label.
        LOG(FATAL) << "Could not find confidence predictions for " << label;
        continue;
      }
      const vector<float>& scores = conf_scores.find(label)->second;
      int loc_label = share_location_ ? -1 : label;
      if (decode_bboxes.find(loc_label) == decode_bboxes.end()) {
        // Something bad happened if there are no predictions for current label.
        LOG(FATAL) << "Could not find location predictions for " << loc_label;
        continue;
      }
      const vector<NormalizedBBox>& bboxes =
          decode_bboxes.find(loc_label)->second;
      vector<int>& indices = it->second;
      if (need_save_) {
        CHECK(label_to_name_.find(label) != label_to_name_.end())
          << "Cannot find label: " << label << " in the label map.";
        CHECK_LT(name_count_, names_.size());
      }
      // 将数据放入输出数据域中
      for (int j = 0; j < indices.size(); ++j) {
        int idx = indices[j];
        top_data[count * 7] = i;
        top_data[count * 7 + 1] = label;
        top_data[count * 7 + 2] = scores[idx];
        const NormalizedBBox& bbox = bboxes[idx];
        top_data[count * 7 + 3] = bbox.xmin();
        top_data[count * 7 + 4] = bbox.ymin();
        top_data[count * 7 + 5] = bbox.xmax();
        top_data[count * 7 + 6] = bbox.ymax();
        if (need_save_) {
          NormalizedBBox out_bbox;
          OutputBBox(bbox, sizes_[name_count_], has_resize_, resize_param_,
                     &out_bbox);
          float score = top_data[count * 7 + 2];
          float xmin = out_bbox.xmin();
          float ymin = out_bbox.ymin();
          float xmax = out_bbox.xmax();
          float ymax = out_bbox.ymax();
          ptree pt_xmin, pt_ymin, pt_width, pt_height;
          pt_xmin.put<float>("", round(xmin * 100) / 100.);
          pt_ymin.put<float>("", round(ymin * 100) / 100.);
          pt_width.put<float>("", round((xmax - xmin) * 100) / 100.);
          pt_height.put<float>("", round((ymax - ymin) * 100) / 100.);

          ptree cur_bbox;
          cur_bbox.push_back(std::make_pair("", pt_xmin));
          cur_bbox.push_back(std::make_pair("", pt_ymin));
          cur_bbox.push_back(std::make_pair("", pt_width));
          cur_bbox.push_back(std::make_pair("", pt_height));

          ptree cur_det;
          cur_det.put("image_id", names_[name_count_]);
          if (output_format_ == "ILSVRC") {
            cur_det.put<int>("category_id", label);
          } else {
            cur_det.put("category_id", label_to_name_[label].c_str());
          }
          cur_det.add_child("bbox", cur_bbox);
          cur_det.put<float>("score", score);

          detections_.push_back(std::make_pair("", cur_det));
        }
        ++count;
      }
    }
    if (need_save_) {
      ++name_count_;
      if (name_count_ % num_test_image_ == 0) {
        if (output_format_ == "VOC") {
          map<string, std::ofstream*> outfiles;
          for (int c = 0; c < num_classes_; ++c) {
            if (c == background_label_id_) {
              continue;
            }
            string label_name = label_to_name_[c];
            boost::filesystem::path file(
                output_name_prefix_ + label_name + ".txt");
            boost::filesystem::path out_file = output_directory / file;
            outfiles[label_name] = new std::ofstream(out_file.string().c_str(),
                std::ofstream::out);
          }
          BOOST_FOREACH(ptree::value_type &det, detections_.get_child("")) {
            ptree pt = det.second;
            string label_name = pt.get<string>("category_id");
            if (outfiles.find(label_name) == outfiles.end()) {
              std::cout << "Cannot find " << label_name << std::endl;
              continue;
            }
            string image_name = pt.get<string>("image_id");
            float score = pt.get<float>("score");
            vector<int> bbox;
            BOOST_FOREACH(ptree::value_type &elem, pt.get_child("bbox")) {
              bbox.push_back(static_cast<int>(elem.second.get_value<float>()));
            }
            *(outfiles[label_name]) << image_name;
            *(outfiles[label_name]) << " " << score;
            *(outfiles[label_name]) << " " << bbox[0] << " " << bbox[1];
            *(outfiles[label_name]) << " " << bbox[0] + bbox[2];
            *(outfiles[label_name]) << " " << bbox[1] + bbox[3];
            *(outfiles[label_name]) << std::endl;
          }
          for (int c = 0; c < num_classes_; ++c) {
            if (c == background_label_id_) {
              continue;
            }
            string label_name = label_to_name_[c];
            outfiles[label_name]->flush();
            outfiles[label_name]->close();
            delete outfiles[label_name];
          }
        } else if (output_format_ == "COCO") {
          boost::filesystem::path output_directory(output_directory_);
          boost::filesystem::path file(output_name_prefix_ + ".json");
          boost::filesystem::path out_file = output_directory / file;
          std::ofstream outfile;
          outfile.open(out_file.string().c_str(), std::ofstream::out);

          boost::regex exp("\"(null|true|false|-?[0-9]+(\\.[0-9]+)?)\"");
          ptree output;
          output.add_child("detections", detections_);
          std::stringstream ss;
          write_json(ss, output);
          std::string rv = boost::regex_replace(ss.str(), exp, "$1");
          outfile << rv.substr(rv.find("["), rv.rfind("]") - rv.find("["))
              << std::endl << "]" << std::endl;
        } else if (output_format_ == "ILSVRC") {
          boost::filesystem::path output_directory(output_directory_);
          boost::filesystem::path file(output_name_prefix_ + ".txt");
          boost::filesystem::path out_file = output_directory / file;
          std::ofstream outfile;
          outfile.open(out_file.string().c_str(), std::ofstream::out);

          BOOST_FOREACH(ptree::value_type &det, detections_.get_child("")) {
            ptree pt = det.second;
            int label = pt.get<int>("category_id");
            string image_name = pt.get<string>("image_id");
            float score = pt.get<float>("score");
            vector<int> bbox;
            BOOST_FOREACH(ptree::value_type &elem, pt.get_child("bbox")) {
              bbox.push_back(static_cast<int>(elem.second.get_value<float>()));
            }
            outfile << image_name << " " << label << " " << score;
            outfile << " " << bbox[0] << " " << bbox[1];
            outfile << " " << bbox[0] + bbox[2];
            outfile << " " << bbox[1] + bbox[3];
            outfile << std::endl;
          }
        }
        name_count_ = 0;
        detections_.clear();
      }
    }
  }
  if (visualize_) {
#ifdef USE_OPENCV
    vector<cv::Mat> cv_imgs;
    this->data_transformer_->TransformInv(bottom[3], &cv_imgs);
    vector<cv::Scalar> colors = GetColors(label_to_display_name_.size());
    VisualizeBBox(cv_imgs, top[0], visualize_threshold_, colors,
        label_to_display_name_, save_file_);
#endif  // USE_OPENCV
  }
}

#ifdef CPU_ONLY
STUB_GPU_FORWARD(DetectionOutputLayer, Forward);
#endif

INSTANTIATE_CLASS(DetectionOutputLayer);
REGISTER_LAYER_CLASS(DetectionOutput);

}  // namespace caffe

caffe定义

message DetectionOutputParameter {
  // 预测种类
  optional uint32 num_classes = 1;
  // 不同类别之间是否共享框位置
  optional bool share_location = 2 [default = true];
  // Background label id. 无则为 -1.
  optional int32 background_label_id = 3 [default = 0];
  // nms参数
  optional NonMaximumSuppressionParameter nms_param = 4;
  // Parameters used for saving detection results.
  optional SaveOutputParameter save_output_param = 5;
  // bbox的编解码方式
  optional PriorBoxParameter.CodeType code_type = 6 [default = CORNER];
  // variance是否被编码
  optional bool variance_encoded_in_target = 8 [default = false];
  // 每张图片在nms处理后保留框的数量
  // -1 表示保留所有框
  optional int32 keep_top_k = 7 [default = -1];
  // 得分阈值
  optional float confidence_threshold = 9;
  // If true, visualize the detection results.
  optional bool visualize = 10 [default = false];
  // The threshold used to visualize the detection results.
  optional float visualize_threshold = 11;
  // If provided, save outputs to video file.
  optional string save_file = 12;
}