301 lines
12 KiB
C++
301 lines
12 KiB
C++
|
||
#include "postprocess.h"
|
||
|
||
|
||
std::tuple<uint8_t, uint8_t, uint8_t> hsv2bgr(float h, float s, float v)
|
||
{
|
||
const int h_i = static_cast<int>(h * 6);
|
||
const float f = h * 6 - h_i;
|
||
const float p = v * (1 - s);
|
||
const float q = v * (1 - f*s);
|
||
const float t = v * (1 - (1 - f) * s);
|
||
float r, g, b;
|
||
switch (h_i) {
|
||
case 0:r = v; g = t; b = p;break;
|
||
case 1:r = q; g = v; b = p;break;
|
||
case 2:r = p; g = v; b = t;break;
|
||
case 3:r = p; g = q; b = v;break;
|
||
case 4:r = t; g = p; b = v;break;
|
||
case 5:r = v; g = p; b = q;break;
|
||
default:r = 1; g = 1; b = 1;break;}
|
||
return std::make_tuple(static_cast<uint8_t>(b * 255), static_cast<uint8_t>(g * 255), static_cast<uint8_t>(r * 255));
|
||
}
|
||
|
||
std::tuple<uint8_t, uint8_t, uint8_t> hsv2rgb(float h, float s, float v)
|
||
{
|
||
const int h_i = static_cast<int>(h * 6);
|
||
const float f = h * 6 - h_i;
|
||
const float p = v * (1 - s);
|
||
const float q = v * (1 - f*s);
|
||
const float t = v * (1 - (1 - f) * s);
|
||
float r, g, b;
|
||
switch (h_i) {
|
||
case 0:r = v; g = t; b = p;break;
|
||
case 1:r = q; g = v; b = p;break;
|
||
case 2:r = p; g = v; b = t;break;
|
||
case 3:r = p; g = q; b = v;break;
|
||
case 4:r = t; g = p; b = v;break;
|
||
case 5:r = v; g = p; b = q;break;
|
||
default:r = 1; g = 1; b = 1;break;}
|
||
return std::make_tuple(static_cast<uint8_t>(r * 255), static_cast<uint8_t>(g * 255), static_cast<uint8_t>(b * 255));
|
||
}
|
||
|
||
std::tuple<uint8_t, uint8_t, uint8_t> randomColor(int id)
|
||
{
|
||
float h_plane = ((((unsigned int)id << 2) ^ 0x937151) % 100) / 100.0f;;
|
||
float s_plane = ((((unsigned int)id << 3) ^ 0x315793) % 100) / 100.0f;
|
||
// return hsv2bgr(h_plane, s_plane, 1);
|
||
return hsv2rgb(h_plane, s_plane, 1);
|
||
}
|
||
|
||
//坐标转换
|
||
void xywh2xyxy(float *xywh, float * xyxy)
|
||
{
|
||
xyxy[0] = (float)(xywh[0] - xywh[2] / 2);
|
||
xyxy[1] = (float)(xywh[1] - xywh[3] / 2);
|
||
xyxy[2] = (float)(xywh[0] + xywh[2] / 2);
|
||
xyxy[3] = (float)(xywh[1] + xywh[3] / 2);
|
||
}
|
||
|
||
//获取区域框1
|
||
cv::Rect getRect(cv::Mat& img, unsigned int uiModelWidth, unsigned int uiModelHeight, float fBbox[4])
|
||
{
|
||
float l, r, t, b;
|
||
float r_w = uiModelWidth / (img.cols * 1.0);
|
||
float r_h = uiModelHeight / (img.rows * 1.0);
|
||
if(r_h > r_w){
|
||
l = fBbox[0] - fBbox[2] / 2.f;
|
||
r = fBbox[0] + fBbox[2] / 2.f;
|
||
t = fBbox[1] - fBbox[3] / 2.f - (uiModelHeight - r_w * img.rows) / 2;
|
||
b = fBbox[1] + fBbox[3] / 2.f - (uiModelHeight - r_w * img.rows) / 2;
|
||
l = l / r_w;
|
||
r = r / r_w;
|
||
t = t / r_w;
|
||
b = b / r_w;
|
||
}else{
|
||
l = fBbox[0] - fBbox[2] / 2.f - (uiModelWidth - r_h * img.cols) / 2;
|
||
r = fBbox[0] + fBbox[2] / 2.f - (uiModelWidth - r_h * img.cols) / 2;
|
||
t = fBbox[1] - fBbox[3] / 2.f;
|
||
b = fBbox[1] + fBbox[3] / 2.f;
|
||
l = l / r_h;
|
||
r = r / r_h;
|
||
t = t / r_h;
|
||
b = b / r_h;
|
||
}
|
||
return cv::Rect(round(l), round(t), round(r - l), round(b - t));
|
||
}
|
||
|
||
//获取区域框2
|
||
cv::Rect getRectangle(cv::Mat& img, float fBbox[4])
|
||
{
|
||
return cv::Rect(round(fBbox[0]), round(fBbox[1]), round(fBbox[2]-fBbox[0]), round(fBbox[3]-fBbox[1]));
|
||
}
|
||
|
||
//计算IOU
|
||
float iou(float fLbox[4], float fRbox[4])
|
||
{
|
||
float interBox[] = {
|
||
(std::max)(fLbox[0] - fLbox[2] / 2.f , fRbox[0] - fRbox[2] / 2.f), //left
|
||
(std::min)(fLbox[0] + fLbox[2] / 2.f , fRbox[0] + fRbox[2] / 2.f), //right
|
||
(std::max)(fLbox[1] - fLbox[3] / 2.f , fRbox[1] - fRbox[3] / 2.f), //top
|
||
(std::min)(fLbox[1] + fLbox[3] / 2.f , fRbox[1] + fRbox[3] / 2.f), //bottom
|
||
};
|
||
|
||
if (interBox[2] > interBox[3] || interBox[0] > interBox[1])
|
||
return 0.0f;
|
||
|
||
float fInterBoxS = (interBox[1] - interBox[0])*(interBox[3] - interBox[2]);
|
||
return fInterBoxS / (fLbox[2] * fLbox[3] + fRbox[2] * fRbox[3] - fInterBoxS);
|
||
}
|
||
|
||
//比较两者置信度
|
||
bool confCmp(const Detection& a, const Detection& b)
|
||
{
|
||
return a.fClassConf > b.fClassConf;
|
||
}
|
||
|
||
//获取缩放比例
|
||
float GetResizeRatio(unsigned int uiImgWidth, unsigned int uiImgHeight, unsigned int uiModelWidth, unsigned int uiModelHeight)
|
||
{
|
||
float fRatioW = static_cast<float>(uiImgWidth) / uiModelWidth;
|
||
float fRatioH = static_cast<float>(uiImgHeight) / uiModelHeight;
|
||
|
||
return (fRatioW - fRatioH > 1e-5) ? fRatioW : fRatioH;
|
||
}
|
||
|
||
//左上顶点补边方式坐标还原
|
||
void UpperVertexResetLocation(float fResizeRatio, unsigned int uiOrigWidth, unsigned int uiOrigHeight, Detection &detection)
|
||
{
|
||
for(int i=0; i<4; i++){
|
||
detection.fBbox[i] = detection.fBbox[i] * fResizeRatio;
|
||
}
|
||
|
||
detection.fBbox[0] = (detection.fBbox[0] < uiOrigWidth) ? detection.fBbox[0] : uiOrigWidth;
|
||
detection.fBbox[1] = (detection.fBbox[1] < uiOrigHeight) ? detection.fBbox[1] : uiOrigHeight;
|
||
detection.fBbox[2] = (detection.fBbox[2] < uiOrigWidth) ? detection.fBbox[2] : uiOrigWidth;
|
||
detection.fBbox[3] = (detection.fBbox[3] < uiOrigHeight) ? detection.fBbox[3] : uiOrigHeight;
|
||
}
|
||
|
||
//中心补边方式坐标还原
|
||
void CenterResetLocation(float fResizeRatio, unsigned int uiOrigWidth, unsigned int uiOrigHeight, unsigned int uiInputWidth, unsigned int uiInputHeight, Detection &detection)
|
||
{
|
||
int w, h, x, y;
|
||
float r_w = uiInputWidth / (uiOrigWidth*1.0);
|
||
float r_h = uiInputHeight / (uiOrigHeight*1.0);
|
||
if (r_h > r_w) {
|
||
w = uiInputWidth;
|
||
h = r_w * uiOrigHeight;
|
||
x = 0;
|
||
y = (uiInputHeight - h) / 2;
|
||
detection.fBbox[1] -= y;
|
||
detection.fBbox[3] -= y;
|
||
} else {
|
||
w = r_h * uiOrigWidth;
|
||
h = uiInputHeight;
|
||
x = (uiInputWidth - w) / 2;
|
||
y = 0;
|
||
detection.fBbox[0] -= x;
|
||
detection.fBbox[2] -= x;
|
||
}
|
||
|
||
for(int i=0; i<4; i++){
|
||
detection.fBbox[i] = detection.fBbox[i] * fResizeRatio;
|
||
}
|
||
|
||
detection.fBbox[0] = (detection.fBbox[0] < uiOrigWidth) ? detection.fBbox[0] : uiOrigWidth;
|
||
detection.fBbox[1] = (detection.fBbox[1] < uiOrigHeight) ? detection.fBbox[1] : uiOrigHeight;
|
||
detection.fBbox[2] = (detection.fBbox[2] < uiOrigWidth) ? detection.fBbox[2] : uiOrigWidth;
|
||
detection.fBbox[3] = (detection.fBbox[3] < uiOrigHeight) ? detection.fBbox[3] : uiOrigHeight;
|
||
}
|
||
|
||
//非极大值抑制,默认0.5
|
||
void yolov5ClearDecodeOpenCVNms(std::vector<ClearDetection>& vecRes, float *fOutput, unsigned int uiOutSize, unsigned int uiDetSize, unsigned int uiClassNum, unsigned int uiClearNum, float fConfThresh = 0.5, float fNmsThresh = 0.4)
|
||
{
|
||
//1.筛选出第一轮结果，根据conf > conf_thresh
|
||
std::vector<int> vecFilterList;
|
||
for(int i = 0; i < uiOutSize /uiDetSize ; ++i){ //锚框数量(遍历锚框筛去置信度过于低的)
|
||
if(fOutput[uiDetSize * i + 4] > fConfThresh){
|
||
vecFilterList.emplace_back(i); // 记录下所有置信度大于conf_thresh的锚框
|
||
}
|
||
}
|
||
if(vecFilterList.size() == 0) return;
|
||
|
||
//2.查找剩余锚框中每个类得分最高值并记录标签
|
||
std::vector<ClearDetection> vecResult;
|
||
std::vector<cv::Rect> vecBoxes;
|
||
std::vector<float> vecScores;
|
||
for(int i : vecFilterList){
|
||
float* pClassConfidence = &fOutput[uiDetSize * i + 5]; //类别
|
||
float fClassConf = *pClassConfidence++;
|
||
int iClassLabel = 0;
|
||
for(int j = 1; j <uiClassNum; ++j, ++pClassConfidence){ //N个类别中查找置信度最高的,并记录标签
|
||
if(*pClassConfidence > fClassConf){
|
||
fClassConf = *pClassConfidence;
|
||
iClassLabel = j;
|
||
}
|
||
}
|
||
|
||
float* pClearConfidence = &fOutput[uiDetSize * i + 5 + uiClassNum]; //清晰度
|
||
float fClearConf = *pClearConfidence++;
|
||
int iClearLabel = 0;
|
||
for(int n = 1; n <uiClearNum; ++n, ++pClearConfidence){ //N个清晰度中查找置信度最高的,并记录标签
|
||
if(*pClearConfidence > fClearConf){
|
||
fClearConf = *pClearConfidence;
|
||
iClearLabel = n;
|
||
}
|
||
}
|
||
|
||
float xywh[4] = {fOutput[i*uiDetSize + 0], fOutput[i*uiDetSize + 1], fOutput[i*uiDetSize + 2], fOutput[i*uiDetSize + 3]};
|
||
float xyxy[4];
|
||
xywh2xyxy(xywh, xyxy); //转换后的坐标
|
||
ClearDetection clearDetection;
|
||
for(int n = 0; n<4; n++){
|
||
clearDetection.detection.fBbox[n] = xyxy[n];
|
||
}
|
||
clearDetection.detection.fClassConf = fOutput[uiDetSize * i + 4] * fClassConf;
|
||
clearDetection.detection.iClassId = iClassLabel;
|
||
clearDetection.fClearConf = fOutput[uiDetSize * i + 4] * fClearConf;
|
||
clearDetection.iClearId = iClearLabel;
|
||
|
||
//总得分大于某阈值才进行NMS,此处阈值可以乘某系数!
|
||
if(clearDetection.detection.fClassConf>fConfThresh) {
|
||
cv::Rect tempRect;
|
||
tempRect.x = xyxy[0];
|
||
tempRect.y = xyxy[1];
|
||
tempRect.width = std::abs(xyxy[2] - xyxy[0]);
|
||
tempRect.height = std::abs(xyxy[3] - xyxy[1]);
|
||
|
||
vecBoxes.push_back(tempRect);
|
||
vecScores.push_back(clearDetection.detection.fClassConf);
|
||
vecResult.push_back(clearDetection);
|
||
}
|
||
}
|
||
|
||
std::vector<int> vecIdx; //保留nms后框的索引
|
||
cv::dnn::NMSBoxes(vecBoxes, vecScores, fConfThresh, fNmsThresh, vecIdx);
|
||
|
||
for (std::size_t i = 0; i < vecIdx.size(); i++){
|
||
vecRes.push_back(vecResult[vecIdx[i]]);
|
||
}
|
||
}
|
||
|
||
|
||
//非极大值抑制,默认0.5
|
||
void yolov5DecodeOpenCVNms(std::vector<Detection>& vecRes, float *fOutput, unsigned int uiOutSize, unsigned int uiDetSize, unsigned int uiClassNum, float fConfThresh = 0.5, float fNmsThresh = 0.4)
|
||
{
|
||
//1.筛选出第一轮结果，根据conf > conf_thresh
|
||
std::vector<int> vecFilterList;
|
||
for(int i = 0; i < uiOutSize /uiDetSize ; ++i){ //锚框数量(遍历锚框筛去置信度过于低的)
|
||
if(fOutput[uiDetSize * i + 4] > fConfThresh){
|
||
vecFilterList.emplace_back(i); // 记录下所有置信度大于conf_thresh的锚框
|
||
}
|
||
}
|
||
if(vecFilterList.size() == 0) return;
|
||
|
||
//2.查找剩余锚框中每个类得分最高值并记录标签
|
||
std::vector<Detection> vecResult;
|
||
std::vector<cv::Rect> vecBoxes;
|
||
std::vector<float> vecScores;
|
||
for(int i : vecFilterList){
|
||
float* pClassConfidence = &fOutput[uiDetSize * i + 5]; //类别
|
||
float fClassConf = *pClassConfidence++;
|
||
int iClassLabel = 0;
|
||
for(int j = 1; j <uiClassNum; ++j, ++pClassConfidence){ //N个类别中查找置信度最高的,并记录标签
|
||
if(*pClassConfidence > fClassConf){
|
||
fClassConf = *pClassConfidence;
|
||
iClassLabel = j;
|
||
}
|
||
}
|
||
|
||
float xywh[4] = {fOutput[i*uiDetSize + 0], fOutput[i*uiDetSize + 1], fOutput[i*uiDetSize + 2], fOutput[i*uiDetSize + 3]};
|
||
float xyxy[4];
|
||
xywh2xyxy(xywh, xyxy); //转换后的坐标
|
||
Detection detection;
|
||
for(int n = 0; n<4; n++){
|
||
detection.fBbox[n] = xyxy[n];
|
||
}
|
||
detection.fClassConf = fOutput[uiDetSize * i + 4] * fClassConf;
|
||
detection.iClassId = iClassLabel;
|
||
|
||
//总得分大于某阈值才进行NMS,此处阈值可以乘某系数!
|
||
if(detection.fClassConf>fConfThresh) {
|
||
cv::Rect tempRect;
|
||
tempRect.x = xyxy[0];
|
||
tempRect.y = xyxy[1];
|
||
tempRect.width = std::abs(xyxy[2] - xyxy[0]);
|
||
tempRect.height = std::abs(xyxy[3] - xyxy[1]);
|
||
|
||
vecBoxes.push_back(tempRect);
|
||
vecScores.push_back(detection.fClassConf);
|
||
vecResult.push_back(detection);
|
||
}
|
||
}
|
||
|
||
std::vector<int> vecIdx; //保留nms后框的索引
|
||
cv::dnn::NMSBoxes(vecBoxes, vecScores, fConfThresh, fNmsThresh, vecIdx);
|
||
|
||
for (std::size_t i = 0; i < vecIdx.size(); i++){
|
||
vecRes.push_back(vecResult[vecIdx[i]]);
|
||
}
|
||
}
|