opencv/modules/objdetect/test/test_face.cpp
Yuantao Feng f77c3574af
Merge pull request #21607 from fengyuentau:fix_FaceDetectorYN_dynamic_shape
Use YuNet of fixed input shape to fix not-supported-dynamic-zero-shape for FaceDetectorYN

* use yunet with input of fixed shape

* update yunet used in face recognition regression
2022-02-21 13:49:07 +00:00

220 lines
7.5 KiB
C++

// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
#include "test_precomp.hpp"
namespace opencv_test { namespace {
// label format:
// image_name
// num_face
// face_1
// face_..
// face_num
std::map<std::string, Mat> blobFromTXT(const std::string& path, int numCoords)
{
std::ifstream ifs(path.c_str());
CV_Assert(ifs.is_open());
std::map<std::string, Mat> gt;
Mat faces;
int faceNum = -1;
int faceCount = 0;
for (std::string line, key; getline(ifs, line); )
{
std::istringstream iss(line);
if (line.find(".png") != std::string::npos)
{
// Get filename
iss >> key;
}
else if (line.find(" ") == std::string::npos)
{
// Get the number of faces
iss >> faceNum;
}
else
{
// Get faces
Mat face(1, numCoords, CV_32FC1);
for (int j = 0; j < numCoords; j++)
{
iss >> face.at<float>(0, j);
}
faces.push_back(face);
faceCount++;
}
if (faceCount == faceNum)
{
// Store faces
gt[key] = faces;
faces.release();
faceNum = -1;
faceCount = 0;
}
}
return gt;
}
TEST(Objdetect_face_detection, regression)
{
// Pre-set params
float scoreThreshold = 0.7f;
float matchThreshold = 0.9f;
float l2disThreshold = 5.0f;
int numLM = 5;
int numCoords = 4 + 2 * numLM;
// Load ground truth labels
std::map<std::string, Mat> gt = blobFromTXT(findDataFile("dnn_face/detection/cascades_labels.txt"), numCoords);
// for (auto item: gt)
// {
// std::cout << item.first << " " << item.second.size() << std::endl;
// }
// Initialize detector
std::string model = findDataFile("dnn/onnx/models/yunet-202202.onnx", false);
Ptr<FaceDetectorYN> faceDetector = FaceDetectorYN::create(model, "", Size(300, 300));
faceDetector->setScoreThreshold(0.7f);
// Detect and match
for (auto item: gt)
{
std::string imagePath = findDataFile("cascadeandhog/images/" + item.first);
Mat image = imread(imagePath);
// Set input size
faceDetector->setInputSize(image.size());
// Run detection
Mat faces;
faceDetector->detect(image, faces);
// std::cout << item.first << " " << item.second.rows << " " << faces.rows << std::endl;
// Match bboxes and landmarks
std::vector<bool> matchedItem(item.second.rows, false);
for (int i = 0; i < faces.rows; i++)
{
if (faces.at<float>(i, numCoords) < scoreThreshold)
continue;
bool boxMatched = false;
std::vector<bool> lmMatched(numLM, false);
cv::Rect2f resBox(faces.at<float>(i, 0), faces.at<float>(i, 1), faces.at<float>(i, 2), faces.at<float>(i, 3));
for (int j = 0; j < item.second.rows && !boxMatched; j++)
{
if (matchedItem[j])
continue;
// Retrieve bbox and compare IoU
cv::Rect2f gtBox(item.second.at<float>(j, 0), item.second.at<float>(j, 1), item.second.at<float>(j, 2), item.second.at<float>(j, 3));
double interArea = (resBox & gtBox).area();
double iou = interArea / (resBox.area() + gtBox.area() - interArea);
if (iou >= matchThreshold)
{
boxMatched = true;
matchedItem[j] = true;
}
// Match landmarks if bbox is matched
if (!boxMatched)
continue;
for (int lmIdx = 0; lmIdx < numLM; lmIdx++)
{
float gtX = item.second.at<float>(j, 4 + 2 * lmIdx);
float gtY = item.second.at<float>(j, 4 + 2 * lmIdx + 1);
float resX = faces.at<float>(i, 4 + 2 * lmIdx);
float resY = faces.at<float>(i, 4 + 2 * lmIdx + 1);
float l2dis = cv::sqrt((gtX - resX) * (gtX - resX) + (gtY - resY) * (gtY - resY));
if (l2dis <= l2disThreshold)
{
lmMatched[lmIdx] = true;
}
}
}
EXPECT_TRUE(boxMatched) << "In image " << item.first << ", cannot match resBox " << resBox << " with any ground truth.";
if (boxMatched)
{
EXPECT_TRUE(std::all_of(lmMatched.begin(), lmMatched.end(), [](bool v) { return v; })) << "In image " << item.first << ", resBox " << resBox << " matched but its landmarks failed to match.";
}
}
}
}
TEST(Objdetect_face_recognition, regression)
{
// Pre-set params
float score_thresh = 0.9f;
float nms_thresh = 0.3f;
double cosine_similar_thresh = 0.363;
double l2norm_similar_thresh = 1.128;
// Load ground truth labels
std::ifstream ifs(findDataFile("dnn_face/recognition/cascades_label.txt").c_str());
CV_Assert(ifs.is_open());
std::set<std::string> fSet;
std::map<std::string, Mat> featureMap;
std::map<std::pair<std::string, std::string>, int> gtMap;
for (std::string line, key; getline(ifs, line);)
{
std::string fname1, fname2;
int label;
std::istringstream iss(line);
iss>>fname1>>fname2>>label;
// std::cout<<fname1<<" "<<fname2<<" "<<label<<std::endl;
fSet.insert(fname1);
fSet.insert(fname2);
gtMap[std::make_pair(fname1, fname2)] = label;
}
// Initialize detector
std::string detect_model = findDataFile("dnn/onnx/models/yunet-202202.onnx", false);
Ptr<FaceDetectorYN> faceDetector = FaceDetectorYN::create(detect_model, "", Size(150, 150), score_thresh, nms_thresh);
std::string recog_model = findDataFile("dnn/onnx/models/face_recognizer_fast.onnx", false);
Ptr<FaceRecognizerSF> faceRecognizer = FaceRecognizerSF::create(recog_model, "");
// Detect and match
for (auto fname: fSet)
{
std::string imagePath = findDataFile("dnn_face/recognition/" + fname);
Mat image = imread(imagePath);
Mat faces;
faceDetector->detect(image, faces);
Mat aligned_face;
faceRecognizer->alignCrop(image, faces.row(0), aligned_face);
Mat feature;
faceRecognizer->feature(aligned_face, feature);
featureMap[fname] = feature.clone();
}
for (auto item: gtMap)
{
Mat feature1 = featureMap[item.first.first];
Mat feature2 = featureMap[item.first.second];
int label = item.second;
double cos_score = faceRecognizer->match(feature1, feature2, FaceRecognizerSF::DisType::FR_COSINE);
double L2_score = faceRecognizer->match(feature1, feature2, FaceRecognizerSF::DisType::FR_NORM_L2);
EXPECT_TRUE(label == 0 ? cos_score <= cosine_similar_thresh : cos_score > cosine_similar_thresh) << "Cosine match result of images " << item.first.first << " and " << item.first.second << " is different from ground truth (score: "<< cos_score <<";Thresh: "<< cosine_similar_thresh <<").";
EXPECT_TRUE(label == 0 ? L2_score > l2norm_similar_thresh : L2_score <= l2norm_similar_thresh) << "L2norm match result of images " << item.first.first << " and " << item.first.second << " is different from ground truth (score: "<< L2_score <<";Thresh: "<< l2norm_similar_thresh <<").";
}
}
}} // namespace