[目标检测]windows下实现c++版faster-rcnn
本文主要参考：http://blog.csdn.net/oYangZi12/article/details/53290426?locationNum=5&fps=1
1.下载微软提供的caffe(https://github.com/Microsoft/caffe)并编译，Pre-Build Steps
Copy .\windows\CommonSettings.props.example to .\windows\CommonSettings.props 选择cpu或者gpu模式。
2.在caffe-master中libcaffe的相应位置加入roi-pooling层对应的cpp，hpp，cu文件重新编译caffe)。因roi-pooling层已经在微软版本的caffe中存在，只需要找到并添加到libcaffe模块下。caffe-master\include\caffe\layers\roi_pooling_layer.hpp，caffe-master\include\caffe\layers\smooth_l1_loss_layer.hpp添加至caffe-master\include\caffe\layers路径下，将caffe-master\src\caffe\layers\roi_pooling_layer.cpp，caffe-master\src\caffe\layers\smooth_l1_loss_layer添加至caffe-master\src\caffe\layers路径下。
3.重新build（生成各模块），并对整个caffe工程进行build.build成功则roi_pooling_layer层添加成功。
4.caffe-master工程下添加Faster_rcnn模块，这里直接将Classification模块中classification.cpp替换掉，Classification重命名成Detect,添加Faster_rcnn.h,Faster_rcnn.cpp和Detect.cpp,添加路径为..\caffe-master\examples\cpp_classification\
5.Faster_rcnn.h代码如下：
#pragma once
#include float> > rpn_net, faster_rcnn_net;
double im_scale;
Size feature_map_size;
private:
bool loadnet();
bool imgtoblob();
vector Faster_rcnn.cpp代码如下：
#include "Faster_rcnn.h"
#include ("faster_rcnn_VOC0712_ZF\\proposal_test.prototxt", caffe::TEST));
rpn_net->CopyTrainedLayersFrom("faster_rcnn_VOC0712_ZF\\proposal_final");
faster_rcnn_net.reset(new Net("faster_rcnn_VOC0712_ZF\\detection_test.prototxt", caffe::TEST));
faster_rcnn_net->CopyTrainedLayersFrom("faster_rcnn_VOC0712_ZF\\detection_final");*/
return true;
}
bool Faster_rcnn::imgtoblob()
{
Mat sample_float;
m_src.convertTo(sample_float, CV_32FC3);
cv::Scalar channel_mean = cv::mean(sample_float);
Mat mean = cv::Mat(m_src.rows, m_src.cols, sample_float.type(), channel_mean);
Mat sample_normalized;
subtract(sample_float, mean, sample_normalized);
prep_im_size();
resize(sample_normalized, sample_normalized, input_geometry_);
Blob<float>* input_layer = rpn_net->input_blobs()[0];
input_layer->Reshape(1, sample_normalized.channels(), sample_normalized.rows, sample_normalized.cols);
rpn_net->Reshape();
float* input_data = input_layer->mutable_cpu_data();
vectorfloat >*>& result = rpn_net->Forward();
Blob<float>* resule0 = result[0];
Blob<float>* resule1 = result[1];
Mat boxs_delta(resule0->num()*resule0->channels()*resule0->width()*resule0->height() / 4, 4, CV_32FC1);
float* p = resule0->mutable_cpu_data();
int num = 0;
for (int i = 0; i < resule0->num()*resule0->channels()*resule0->width()*resule0->height() / 4; i++)
{
for (int j = 0; j < 4; j++)
{
boxs_delta.at<float>(i, j) = resule0->data_at(0, num%resule0->channels(),
(num - num / resule0->channels() / resule0->height() * resule0->channels() * resule0->height()) / resule0->height(),
num / resule0->channels() / resule0->height());
num++;
//int order = j + i * 4;
//boxs_delta.at(i, j) = resule0->data_at(0, (order % (resule0->height()*resule0->channels())) % resule0->channels(), (order % (resule0->height()*resule0->channels())) / resule0->channels(), order / (resule0->height()*resule0->channels()));
}
}
//create anchors
feature_map_size = Size(resule0->width(), resule0->height());
//prep_im_size();
Mat anchors = proposal_local_anchor();
Mat pre_box = bbox_tranform_inv(anchors, boxs_delta, "rpn");
//Mat score(resule0->width(), resule0->height(), CV_32FC1);
Mat score = get_rpn_score(resule1, resule0->width(), resule0->height());
vector tmp;
int channel = resule1->width()*resule1->height() / (w * h);
Mat score(resule1->width()*resule1->height(), 1, CV_32FC1);
//tmp.Reshape(1, resule1->width()*resule1->height() / (w * h),h,w);
//float* p = tmp.mutable_cpu_data;
int num = 0;
for (int i = 0; i < w; i++)
{
for (int j = 0; j < h; j++)
{
for (int k = 0; k < channel; k++)
{
score.at<float>(num, 0) = resule1->data_at(0, 1, h*k + j, i);
num++;
}
}
}
return score;
}
Mat Faster_rcnn::bbox_tranform_inv(Mat anchors, Mat boxs_delta, string type)
{
if (type == "rpn")
{
Mat pre_box(anchors.rows, anchors.cols, CV_32FC1);
for (int i = 0; i < anchors.rows; i++)
{
double pred_ctr_x, pred_ctr_y, src_ctr_x, src_ctr_y;
double dst_ctr_x, dst_ctr_y, dst_scl_x, dst_scl_y;
double src_w, src_h, pred_w, pred_h;
src_w = anchors.at<float>(i, 2) - anchors.at<float>(i, 0) + 1;
src_h = anchors.at<float>(i, 3) - anchors.at<float>(i, 1) + 1;
src_ctr_x = anchors.at<float>(i, 0) + 0.5 * (src_w - 1);
src_ctr_y = anchors.at<float>(i, 1) + 0.5 * (src_h - 1);
dst_ctr_x = boxs_delta.at<float>(i, 0);
dst_ctr_y = boxs_delta.at<float>(i, 1);
dst_scl_x = boxs_delta.at<float>(i, 2);
dst_scl_y = boxs_delta.at<float>(i, 3);
pred_ctr_x = dst_ctr_x*src_w + src_ctr_x;
pred_ctr_y = dst_ctr_y*src_h + src_ctr_y;
pred_w = exp(dst_scl_x) * src_w;
pred_h = exp(dst_scl_y) * src_h;
pre_box.at<float>(i, 0) = ((pred_ctr_x - 0.5*(pred_w - 1)) - 1) * (float)(m_src.cols) / (im.cols - 1) + 1;
pre_box.at<float>(i, 1) = ((pred_ctr_y - 0.5*(pred_h - 1)) - 1) * (float)(m_src.rows) / (im.rows - 1) + 1;
pre_box.at<float>(i, 2) = ((pred_ctr_x + 0.5*(pred_w - 1)) - 1) * (float)(m_src.cols) / (im.cols - 1) + 1;
pre_box.at<float>(i, 3) = ((pred_ctr_y + 0.5*(pred_h - 1)) - 1) * (float)(m_src.rows) / (im.rows - 1) + 1;
if (pre_box.at<float>(i, 0) < 0) pre_box.at<float>(i, 0) = 0;
if (pre_box.at<float>(i, 0) > (m_src.cols - 1)) pre_box.at<float>(i, 0) = m_src.cols - 1;
if (pre_box.at<float>(i, 2) < 0) pre_box.at<float>(i, 2) = 0;
if (pre_box.at<float>(i, 2) > (m_src.cols - 1)) pre_box.at<float>(i, 2) = m_src.cols - 1;
if (pre_box.at<float>(i, 1) < 0) pre_box.at<float>(i, 1) = 0;
if (pre_box.at<float>(i, 1) > (m_src.rows - 1)) pre_box.at<float>(i, 1) = m_src.rows - 1;
if (pre_box.at<float>(i, 3) < 0) pre_box.at<float>(i, 3) = 0;
if (pre_box.at<float>(i, 3) > (m_src.rows - 1)) pre_box.at<float>(i, 3) = m_src.rows - 1;
}
return pre_box;
}
if (type == "rcnn")
{
Mat pre_box(boxs_delta.rows, boxs_delta.cols, CV_32FC1);
for (int i = 0; i < boxs_delta.rows; i++)
{
for (int j = 1; j < boxs_delta.cols / 4; j++)
{
double pred_ctr_x, pred_ctr_y, src_ctr_x, src_ctr_y;
double dst_ctr_x, dst_ctr_y, dst_scl_x, dst_scl_y;
double src_w, src_h, pred_w, pred_h;
src_w = anchors.at<float>(i, 2) - anchors.at<float>(i, 0) + 1;
src_h = anchors.at<float>(i, 3) - anchors.at<float>(i, 1) + 1;
src_ctr_x = anchors.at<float>(i, 0) + 0.5 * (src_w - 1);
src_ctr_y = anchors.at<float>(i, 1) + 0.5 * (src_h - 1);
dst_ctr_x = boxs_delta.at<float>(i, 4 * j + 0);
dst_ctr_y = boxs_delta.at<float>(i, 4 * j + 1);
dst_scl_x = boxs_delta.at<float>(i, 4 * j + 2);
dst_scl_y = boxs_delta.at<float>(i, 4 * j + 3);
pred_ctr_x = dst_ctr_x*src_w + src_ctr_x;
pred_ctr_y = dst_ctr_y*src_h + src_ctr_y;
pred_w = exp(dst_scl_x) * src_w;
pred_h = exp(dst_scl_y) * src_h;
pre_box.at<float>(i, 4 * (j - 1) + 0) = ((pred_ctr_x - 0.5*(pred_w - 1)) - 1);
pre_box.at<float>(i, 4 * (j - 1) + 1) = ((pred_ctr_y - 0.5*(pred_h - 1)) - 1);
pre_box.at<float>(i, 4 * (j - 1) + 2) = ((pred_ctr_x + 0.5*(pred_w - 1)) - 1);
pre_box.at<float>(i, 4 * (j - 1) + 3) = ((pred_ctr_y + 0.5*(pred_h - 1)) - 1);
if (pre_box.at<float>(i, 4 * (j - 1) + 0) < 0) pre_box.at<float>(i, 4 * (j - 1) + 0) = 0;
if (pre_box.at<float>(i, 4 * (j - 1) + 0) > (m_src.cols - 1)) pre_box.at<float>(i, 4 * (j - 1) + 0) = m_src.cols - 1;
if (pre_box.at<float>(i, 4 * (j - 1) + 2) < 0) pre_box.at<float>(i, 4 * (j - 1) + 2) = 0;
if (pre_box.at<float>(i, 4 * (j - 1) + 2) > (m_src.cols - 1)) pre_box.at<float>(i, 4 * (j - 1) + 2) = m_src.cols - 1;
if (pre_box.at<float>(i, 4 * (j - 1) + 1) < 0) pre_box.at<float>(i, 4 * (j - 1) + 1) = 0;
if (pre_box.at<float>(i, 4 * (j - 1) + 1) > (m_src.rows - 1)) pre_box.at<float>(i, 4 * (j - 1) + 1) = m_src.rows - 1;
if (pre_box.at<float>(i, 4 * (j - 1) + 3) < 0) pre_box.at<float>(i, 4 * (j - 1) + 3) = 0;
if (pre_box.at<float>(i, 4 * (j - 1) + 3) > (m_src.rows - 1)) pre_box.at<float>(i, 4 * (j - 1) + 3) = m_src.rows - 1;
}
}
return pre_box;
}
}
Mat Faster_rcnn::proposal_local_anchor()
{
Blob<float> shift;
Mat shitf_x(feature_map_size.height, feature_map_size.width, CV_32SC1);
Mat shitf_y(feature_map_size.height, feature_map_size.width, CV_32SC1);
for (size_t i = 0; i < feature_map_size.width; i++)
{
for (size_t j = 0; j < feature_map_size.height; j++)
{
shitf_x.at<int>(j, i) = i * conf.feat_stride;
shitf_y.at<int>(j, i) = j * conf.feat_stride;
}
}
shift.Reshape(9, feature_map_size.width*feature_map_size.height, 4, 1);
float *p = shift.mutable_cpu_diff(), *a = shift.mutable_cpu_data();
for (int i = 0; i < feature_map_size.width*feature_map_size.height; i++)
{
for (int j = 0; j < 9; j++)
{
{
size_t num = i * 4 + j * 4 * feature_map_size.width*feature_map_size.height;
p[num + 0] = -shitf_x.at<int>(i % shitf_x.rows, i / shitf_x.rows);
p[num + 2] = -shitf_x.at<int>(i % shitf_x.rows, i / shitf_x.rows);
p[num + 1] = -shitf_y.at<int>(i % shitf_y.rows, i / shitf_y.rows);
p[num + 3] = -shitf_y.at<int>(i % shitf_y.rows, i / shitf_y.rows);
a[num + 0] = conf.anchor[j][0];
a[num + 1] = conf.anchor[j][1];
a[num + 2] = conf.anchor[j][2];
a[num + 3] = conf.anchor[j][3];
}
}
}
shift.Update();
Mat anchors(9 * feature_map_size.width*feature_map_size.height, 4, CV_32FC1);
size_t num = 0;
for (int i = 0; i < anchors.cols; i++)
{
for (int j = 0; j < anchors.rows; j++)
{
anchors.at<float>(j, i) = shift.data_at(num%shift.num(),
(num - num / (shift.num() * shift.channels())*shift.num() * shift.channels()) / shift.num(),
num / (shift.num() * shift.channels()), 0);
num++;
}
}
/*for (int i = 0; i < 4; i++)
{
for (int k = 0; k < feature_map_size.width*feature_map_size.height; k++)
{
for (int j = 0; j < 9; j++)
{
anchors.at(num%anchors.rows, num / anchors.rows) = shift.data_at(j, k, i, 0);
num++;
}
}
}*/
return anchors;
}
void Faster_rcnn::prep_im_size()
{
int im_size_min = min(m_src.cols, m_src.rows);
int im_size_max = max(m_src.cols, m_src.rows);
im_scale = double(conf.target_size) / im_size_min;
if (round(im_scale*im_size_max) > conf.maxsize)
{
im_scale = double(conf.maxsize) / im_size_max;
}
input_geometry_ = Size(round(m_src.cols * im_scale), round(m_src.rows * im_scale));
resize(m_src, im, input_geometry_);
}
bool Faster_rcnn::rpn_converttoboxs()
{
return true;
}
Mat Faster_rcnn::gettarget(Mat src)
{
m_src = src;
//img=>blob
imgtoblob();
vector a = rpn_net->blob_names();
// int b = rpn_net->blob_by_name("conv5")->count();
faster_rcnn_net->blob_by_name("data")->CopyFrom(*rpn_net->blob_by_name("conv5").get(), false, true);
Blob<float>* input_layer = faster_rcnn_net->input_blobs()[1];
int sub_blob_num = aboxes.size();
input_layer->Reshape(sub_blob_num, 5, 1, 1);
float* input_data = input_layer->mutable_cpu_data();
int num = 0;
for (auto ite = aboxes.begin(); ite != aboxes.end(); ite++)
{
anchors.at<float>(num, 0) = ite->x1;
anchors.at<float>(num, 1) = ite->y1;
anchors.at<float>(num, 2) = ite->x2;
anchors.at<float>(num, 3) = ite->y2;
ite->score = 0;
ite->x1 = (ite->x1 - 1) * scales + 1 - 1;
ite->y1 = (ite->y1 - 1) * scales + 1 - 1;
ite->x2 = (ite->x2 - 1) * scales + 1 - 1;
ite->y2 = (ite->y2 - 1) * scales + 1 - 1;
input_data[num * 5 + 0] = ite->score;
input_data[num * 5 + 1] = ite->x1;
input_data[num * 5 + 2] = ite->y1;
input_data[num * 5 + 3] = ite->x2;
input_data[num * 5 + 4] = ite->y2;
num++;
}
const vectorfloat >*>& result = faster_rcnn_net->Forward();
Blob<float>* scores = result[1];
Blob<float>* box_deltas = result[0];
box_deltas->data_at(0, 3, 0, 0);
Mat boxs_delta(box_deltas->num(), box_deltas->channels(), CV_32FC1);
for (int i = 0; i < box_deltas->num(); i++)
{
for (int j = 0; j < box_deltas->channels(); j++)
{
boxs_delta.at<float>(i, j) = box_deltas->data_at(i, j, 0, 0);
}
}
Mat score(scores->num(), scores->channels() - 1, CV_32FC1);
for (int i = 0; i < scores->num(); i++)
{
for (int j = 0 + 1; j < scores->channels(); j++)
{
score.at<float>(i, j - 1) = scores->data_at(i, j, 0, 0);
}
}
Mat pred_boxes = bbox_tranform_inv(anchors, boxs_delta, "rcnn");
std::map<int, vectorDetect.cpp代码如下：
#include "Faster_rcnn.h"
//#include 6.将matlab版faster-rcnn的ZF网络模型文件放到工程faster_rcnn_VOC0712_ZF文件夹下。https://github.com/ShaoqingRen/faster_rcnn 最后几行给出了下载地址，下载地址为：https://pan.baidu.com/s/1hsFKmeK，下载完成放在自己的路径下，并在Faster_rcnn.cpp文件bool Faster_rcnn::loadnet()函数中正确添加所需的模型文件路径。
7. build(生成）Detect模块，如果是Debug模式，则可以在….\caffe-master\Build\x64\Debug\路径下找到Detect.exe，
cmd进入到….\caffe-master\Build\x64\Debug\路径，执行
..\x64\Debug>Detect.exe
8.出来检测结果：
因图片大小问题不能上传成功。