Skip to content

Commit 2c215ad

Browse files
zhanglaplacezhanglaplace
committed
first commit
1 parent ffbe11d commit 2c215ad

17 files changed

+1040
-1
lines changed

‎CMakeLists.txt‎

Lines changed: 17 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,17 @@
1+
cmake_minimum_required(VERSION 2.9)
2+
project(faster_rcnn)
3+
4+
#set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
5+
set(CMAKE_CXX_STANDARD 11)
6+
7+
find_package(OpenCV)
8+
9+
find_package(Caffe)
10+
#message(FATAL_ERROR ${Caffe_INCLUDE_DIRS})
11+
include_directories(${Caffe_INCLUDE_DIRS})
12+
13+
set(SOURCE_FILES demo.cpp faster-api.cpp faster-api.h)
14+
add_executable(faster_rcnn ${SOURCE_FILES})
15+
16+
target_link_libraries(faster_rcnn ${OpenCV_LIBS} )
17+
target_link_libraries(faster_rcnn ${Caffe_LIBRARIES})

‎README.md‎

Lines changed: 106 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,2 +1,107 @@
11
# faster-rcnn-cplusplus2
2-
faster-rcnn c++ pythonmodel
2+
3+
Faster-rcnn cplusplus2 with python model ;
4+
5+
the project is according to the http://blog.csdn.net/zxj942405301/article/details/72775463
6+
7+
the c++ faster-rcnn with matlab model see:https://github.com/zhanglaplace/Faster_rcnn_Cplusplus_vs2013
8+
# Platform
9+
10+
## Windows
11+
12+
You need a VC compiler to build these project, Visual Studio 2013 Community should be fine. You can download from https://www.visualstudio.com/downloads/.
13+
14+
## Ubuntu
15+
mkdir build
16+
17+
cd build
18+
19+
cmake ..
20+
21+
make .
22+
23+
of course ,you shoule put your model and protoxt file in models directory
24+
25+
# Caffe
26+
You should build caffe with RoiPooling layer and rpn_layer
27+
28+
- rpn_layer
29+
30+
-- add rpn_layer.cpp to $Caffe/src/caffe/layers/
31+
32+
-- add rpn_layer.hpp to $Caffe/include/caffe/layers/
33+
34+
- caffe.proto
35+
```cpp
36+
optional RPNParameter rpn_param = 158;
37+
message RPNParameter {
38+
optional uint32 feat_stride = 1;
39+
optional uint32 basesize = 2;
40+
repeated uint32 scale = 3;
41+
repeated float ratio = 4;
42+
optional uint32 boxminsize =5;
43+
optional uint32 per_nms_topn = 9;
44+
optional uint32 post_nms_topn = 11;
45+
optional float nms_thresh = 8;
46+
}
47+
```
48+
49+
- protoxt modify
50+
```cpp
51+
modify test.prototxt
52+
layer {
53+
name: 'proposal'
54+
type: 'Python'
55+
bottom: 'rpn_cls_prob_reshape'
56+
bottom: 'rpn_bbox_pred'
57+
bottom: 'im_info'
58+
top: 'rois'
59+
python_param {
60+
module: 'rpn.proposal_layer'
61+
layer: 'ProposalLayer'
62+
param_str: "'feat_stride': 16"
63+
}
64+
}
65+
to:
66+
layer {
67+
name: "proposal"
68+
type: "RPN"
69+
bottom: "rpn_cls_prob_reshape"
70+
bottom: "rpn_bbox_pred"
71+
bottom: "im_info"
72+
top: "rois"
73+
rpn_param {
74+
feat_stride : 16
75+
basesize : 16
76+
scale : 8
77+
scale : 16
78+
scale : 32
79+
ratio : 0.5
80+
ratio : 1
81+
ratio : 2
82+
boxminsize :16
83+
per_nms_topn : 0;
84+
post_nms_topn : 0;
85+
nms_thresh : 0.3
86+
}
87+
}
88+
```
89+
90+
# Result
91+
it's cost much time the py-faster-rcnn
92+
In My compute(GTX1080Ti GPU) a picture of size(375\*500\*3)cost 361ms .
93+
94+
![image](https://github.com/zhanglaplace/Faster_rcnn_Cplusplus_vs2013/blob/master/imgs/result_004545.jpg)
95+
![image](https://github.com/zhanglaplace/Faster_rcnn_Cplusplus_vs2013/blob/master/imgs/result_001150.jpg)
96+
![image](https://github.com/zhanglaplace/Faster_rcnn_Cplusplus_vs2013/blob/master/imgs/result_000456.jpg)
97+
![image](https://github.com/zhanglaplace/Faster_rcnn_Cplusplus_vs2013/blob/master/imgs/result_000542.jpg)
98+
![image](https://github.com/zhanglaplace/Faster_rcnn_Cplusplus_vs2013/blob/master/imgs/result_001763.jpg)
99+
100+
101+
# Something else
102+
103+
if it's helpful to you ,please give me a star thanks~
104+
105+
106+
107+

‎demo.cpp‎

Lines changed: 52 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,52 @@
1+
#include "faster-api.h"
2+
#include<opencv2/opencv.hpp>
3+
#include <chrono>
4+
#include<iostream>
5+
#include<sstream>
6+
using namespace cv;
7+
using namespace std;
8+
string num2str(float i){
9+
stringstream ss;
10+
ss<<i;
11+
return ss.str();
12+
}
13+
14+
int main(int argc,char **argv){
15+
::google::InitGoogleLogging(argv[0]);
16+
#ifdef CPU_ONLY
17+
cout<<"Use CPU\n";
18+
#else
19+
cout<<"Use GPU\n";
20+
#endif
21+
22+
ObjectDetector detect("../models/test.prototxt","../models/VGG16_faster_rcnn_final.caffemodel");
23+
ifstream fin("../testImage.txt",std::ios_base::in);
24+
string lines;
25+
while(getline(fin,lines)){
26+
cout<<lines<<endl;
27+
std::chrono::time_point<std::chrono::system_clock> p0 = std::chrono::system_clock::now();
28+
Mat img = imread(lines.c_str());
29+
if(img.empty()){
30+
cout<<"input image error: "<<lines<<endl;
31+
continue;
32+
}
33+
map<int,vector<float> > score;
34+
map<int,vector<Rect> > label_objs=detect.detect(img,&score); //目标检测,同时保存每个框的置信度
35+
36+
for(map<int,vector<Rect> >::iterator it=label_objs.begin();it!=label_objs.end();it++){
37+
int label=it->first; //标签
38+
vector<Rect> rects=it->second; //检测框
39+
for(int j=0;j<rects.size();j++){
40+
rectangle(img,rects[j],Scalar(0,0,255),2); //画出矩形框
41+
string txt=num2str(label)+" : "+num2str(score[label][j]);
42+
putText(img,txt,Point(rects[j].x,rects[j].y),CV_FONT_HERSHEY_SIMPLEX,0.5,Scalar(0,255,0)); //标记 类别：置信度
43+
}
44+
}
45+
std::chrono::time_point<std::chrono::system_clock> p1 = std::chrono::system_clock::now();
46+
cout << "detection time:" << (float)std::chrono::duration_cast<std::chrono::microseconds>(p1 - p0).count() / 1000 << "ms" << endl;
47+
imshow("", img);
48+
waitKey();
49+
}
50+
fin.close();
51+
return 0;
52+
}

‎faster-api.cpp‎

Lines changed: 202 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,202 @@
1+
#include "faster-api.h"
2+
#include <opencv2/highgui/highgui.hpp>
3+
#include <opencv2/imgproc/imgproc.hpp>
4+
#include <vector>
5+
#include <fstream>
6+
7+
using std::string;
8+
using std::vector;
9+
using namespace caffe;
10+
using std::max;
11+
using std::min;
12+
13+
14+
15+
cv::Mat bbox_tranform_inv(cv::Mat local_anchors, cv::Mat boxs_delta){
16+
cv::Mat pre_box(local_anchors.rows, local_anchors.cols, CV_32FC1);
17+
for (int i = 0; i < local_anchors.rows; i++)
18+
{
19+
double pred_ctr_x, pred_ctr_y, src_ctr_x, src_ctr_y;
20+
double dst_ctr_x, dst_ctr_y, dst_scl_x, dst_scl_y;
21+
double src_w, src_h, pred_w, pred_h;
22+
src_w = local_anchors.at<float>(i, 2) - local_anchors.at<float>(i, 0) + 1;
23+
src_h = local_anchors.at<float>(i, 3) - local_anchors.at<float>(i, 1) + 1;
24+
src_ctr_x = local_anchors.at<float>(i, 0) + 0.5 * src_w;
25+
src_ctr_y = local_anchors.at<float>(i, 1) + 0.5 * src_h;
26+
27+
dst_ctr_x = boxs_delta.at<float>(i, 0);
28+
dst_ctr_y = boxs_delta.at<float>(i, 1);
29+
dst_scl_x = boxs_delta.at<float>(i, 2);
30+
dst_scl_y = boxs_delta.at<float>(i, 3);
31+
pred_ctr_x = dst_ctr_x*src_w + src_ctr_x;
32+
pred_ctr_y = dst_ctr_y*src_h + src_ctr_y;
33+
pred_w = exp(dst_scl_x) * src_w;
34+
pred_h = exp(dst_scl_y) * src_h;
35+
36+
pre_box.at<float>(i, 0) = pred_ctr_x - 0.5*pred_w;
37+
pre_box.at<float>(i, 1) = pred_ctr_y - 0.5*pred_h;
38+
pre_box.at<float>(i, 2) = pred_ctr_x + 0.5*pred_w;
39+
pre_box.at<float>(i, 3) = pred_ctr_y + 0.5*pred_h;
40+
}
41+
return pre_box;
42+
}
43+
44+
void nms(std::vector<abox> &input_boxes, float nms_thresh){
45+
std::vector<float>vArea(input_boxes.size());
46+
for (int i = 0; i < input_boxes.size(); ++i)
47+
{
48+
vArea[i] = (input_boxes.at(i).x2 - input_boxes.at(i).x1 + 1)
49+
* (input_boxes.at(i).y2 - input_boxes.at(i).y1 + 1);
50+
}
51+
for (int i = 0; i < input_boxes.size(); ++i)
52+
{
53+
for (int j = i + 1; j < input_boxes.size();)
54+
{
55+
float xx1 = std::max(input_boxes[i].x1, input_boxes[j].x1);
56+
float yy1 = std::max(input_boxes[i].y1, input_boxes[j].y1);
57+
float xx2 = std::min(input_boxes[i].x2, input_boxes[j].x2);
58+
float yy2 = std::min(input_boxes[i].y2, input_boxes[j].y2);
59+
float w = std::max(float(0), xx2 - xx1 + 1);
60+
float h = std::max(float(0), yy2 - yy1 + 1);
61+
float inter = w * h;
62+
float ovr = inter / (vArea[i] + vArea[j] - inter);
63+
if (ovr >= nms_thresh)
64+
{
65+
input_boxes.erase(input_boxes.begin() + j);
66+
vArea.erase(vArea.begin() + j);
67+
}
68+
else
69+
{
70+
j++;
71+
}
72+
}
73+
}
74+
}
75+
76+
77+
78+
79+
ObjectDetector::ObjectDetector(const std::string &model_file,const std::string &weights_file){
80+
#ifdef CPU_ONLY
81+
Caffe::set_mode(Caffe::CPU);
82+
#else
83+
Caffe::set_mode(Caffe::GPU);
84+
#endif
85+
net_.reset(new Net<float>(model_file, TEST));
86+
net_->CopyTrainedLayersFrom(weights_file);
87+
this->class_num_ = net_->blob_by_name("cls_prob")->channels(); //求得类别数+1
88+
}
89+
90+
//对一张图片，进行检测，将结果保存进map数据结构里,分别表示每个类别对应的目标框，如果需要分数信息，则计算分数
91+
map<int,vector<cv::Rect> > ObjectDetector::detect(const cv::Mat& image,map<int,vector<float> >* objectScore){
92+
93+
if(objectScore!=NULL) //如果需要保存置信度
94+
objectScore->clear();
95+
96+
float CONF_THRESH = 0.8; //置信度阈值
97+
float NMS_THRESH = 0.3; //非极大值抑制阈值
98+
int max_side = max(image.rows, image.cols); //分别求出图片宽和高的较大者
99+
int min_side = min(image.rows, image.cols);
100+
float max_side_scale = float(max_side) / float(INPUT_SIZE_LONG); //分别求出缩放因子
101+
float min_side_scale = float(min_side) / float(INPUT_SIZE_NARROW);
102+
float max_scale = max(max_side_scale, min_side_scale);
103+
104+
float img_scale = float(1) / max_scale;
105+
int height = int(image.rows * img_scale);
106+
int width = int(image.cols * img_scale);
107+
108+
int num_out;
109+
cv::Mat cv_resized;
110+
image.convertTo(cv_resized, CV_32FC3);
111+
cv::resize(cv_resized, cv_resized, cv::Size(width, height));
112+
cv::Mat mean(height, width, cv_resized.type(), cv::Scalar(102.9801, 115.9465, 122.7717));
113+
cv::Mat normalized;
114+
subtract(cv_resized, mean, normalized);
115+
116+
float im_info[3];
117+
im_info[0] = height;
118+
im_info[1] = width;
119+
im_info[2] = img_scale;
120+
boost::shared_ptr<Blob<float> > input_layer = net_->blob_by_name("data");
121+
input_layer->Reshape(1, normalized.channels(), height, width);
122+
net_->Reshape();
123+
float* input_data = input_layer->mutable_cpu_data();
124+
vector<cv::Mat> input_channels;
125+
for (int i = 0; i < input_layer->channels(); ++i) {
126+
cv::Mat channel(height, width, CV_32FC1, input_data);
127+
input_channels.push_back(channel);
128+
input_data += height * width;
129+
}
130+
cv::split(normalized, input_channels);
131+
net_->blob_by_name("im_info")->set_cpu_data(im_info);
132+
net_->Forward(); //进行网络前向传播
133+
134+
135+
int num = net_->blob_by_name("rois")->num(); //产生的 ROI 个数,比如为 13949个ROI
136+
const float *rois_data = net_->blob_by_name("rois")->cpu_data(); //维度比如为：13949*5*1*1
137+
int num1 = net_->blob_by_name("bbox_pred")->num(); //预测的矩形框 维度为 13949*84
138+
cv::Mat rois_box(num, 4, CV_32FC1);
139+
for (int i = 0; i < num; ++i)
140+
{
141+
rois_box.at<float>(i, 0) = rois_data[i * 5 + 1] / img_scale;
142+
rois_box.at<float>(i, 1) = rois_data[i * 5 + 2] / img_scale;
143+
rois_box.at<float>(i, 2) = rois_data[i * 5 + 3] / img_scale;
144+
rois_box.at<float>(i, 3) = rois_data[i * 5 + 4] / img_scale;
145+
}
146+
147+
boost::shared_ptr<Blob<float> > bbox_delt_data = net_->blob_by_name("bbox_pred"); // 13949*84
148+
boost::shared_ptr<Blob<float> > score = net_->blob_by_name("cls_prob"); // 3949*21
149+
150+
map<int,vector<cv::Rect> > label_objs; //每个类别，对应的检测目标框
151+
for (int i = 1; i < class_num_; ++i){ //对每个类，进行遍历
152+
cv::Mat bbox_delt(num, 4, CV_32FC1);
153+
for (int j = 0; j < num; ++j){
154+
bbox_delt.at<float>(j, 0) = bbox_delt_data->data_at(j, i * 4 + 0, 0, 0);
155+
bbox_delt.at<float>(j, 1) = bbox_delt_data->data_at(j, i * 4 + 1, 0, 0);
156+
bbox_delt.at<float>(j, 2) = bbox_delt_data->data_at(j, i * 4 + 2, 0, 0);
157+
bbox_delt.at<float>(j, 3) = bbox_delt_data->data_at(j, i * 4 + 3, 0, 0);
158+
}
159+
cv::Mat box_class = bbox_tranform_inv(rois_box, bbox_delt);
160+
161+
vector<abox> aboxes; //对于 类别i，检测出的矩形框保存在这
162+
for (int j = 0; j < box_class.rows; ++j){
163+
if (box_class.at<float>(j, 0) < 0) box_class.at<float>(j, 0) = 0;
164+
if (box_class.at<float>(j, 0) > (image.cols - 1)) box_class.at<float>(j, 0) = image.cols - 1;
165+
if (box_class.at<float>(j, 2) < 0) box_class.at<float>(j, 2) = 0;
166+
if (box_class.at<float>(j, 2) > (image.cols - 1)) box_class.at<float>(j, 2) = image.cols - 1;
167+
168+
if (box_class.at<float>(j, 1) < 0) box_class.at<float>(j, 1) = 0;
169+
if (box_class.at<float>(j, 1) > (image.rows - 1)) box_class.at<float>(j, 1) = image.rows - 1;
170+
if (box_class.at<float>(j, 3) < 0) box_class.at<float>(j, 3) = 0;
171+
if (box_class.at<float>(j, 3) > (image.rows - 1)) box_class.at<float>(j, 3) = image.rows - 1;
172+
abox tmp;
173+
tmp.x1 = box_class.at<float>(j, 0);
174+
tmp.y1 = box_class.at<float>(j, 1);
175+
tmp.x2 = box_class.at<float>(j, 2);
176+
tmp.y2 = box_class.at<float>(j, 3);
177+
tmp.score = score->data_at(j, i, 0, 0);
178+
aboxes.push_back(tmp);
179+
}
180+
std::sort(aboxes.rbegin(), aboxes.rend());
181+
nms(aboxes, NMS_THRESH); //与非极大值抑制消除对于的矩形框
182+
for (int k = 0; k < aboxes.size();){
183+
if (aboxes[k].score < CONF_THRESH)
184+
aboxes.erase(aboxes.begin() + k);
185+
else
186+
k++;
187+
}
188+
//################ 将类别i的所有检测框，保存
189+
vector<cv::Rect> rect(aboxes.size()); //对于类别i，检测出的矩形框
190+
for(int ii=0;ii<aboxes.size();++ii)
191+
rect[ii]=cv::Rect(cv::Point(aboxes[ii].x1,aboxes[ii].y1),cv::Point(aboxes[ii].x2,aboxes[ii].y2));
192+
label_objs[i]=rect;
193+
//################ 将类别i的所有检测框的打分，保存
194+
if(objectScore!=NULL){ //################ 将类别i的所有检测框的打分，保存
195+
vector<float> tmp(aboxes.size()); //对于 类别i，检测出的矩形框的得分
196+
for(int ii=0;ii<aboxes.size();++ii)
197+
tmp[ii]=aboxes[ii].score;
198+
objectScore->insert(pair<int,vector<float> >(i,tmp));
199+
}
200+
}
201+
return label_objs;
202+
}

0 commit comments

Comments
 (0)