opencv/modules/features2d/test/test_detectors_invariance.impl.hpp
Maksim Shabunin f9e9567870
Merge pull request #25424 from mshabunin:fix-features2d-test
features2d: reduce accuracy test avg memory consumption #25424

**Merge with contrib**: https://github.com/opencv/opencv_contrib/pull/3722

I've observed high memory consumption in the `opencv_test_features2d` (x86_64, Ubuntu 22.04, Debug):

![image](https://github.com/opencv/opencv/assets/3304494/419d65d9-d727-4d1e-bdec-dbde6681c188)
It's always more than 180 MiB with peak at 535 MiB

This was caused by pointers to the algorithm object instances stored in the tests parameters. I've replaced them with factory functions/lambdas with the following result:

![image](https://github.com/opencv/opencv/assets/3304494/bd4ff0ea-3db4-4ab8-8e6d-192a3826e99c)
Now peak is at 355 MiB and permanent consumption level is ~ 1-2 MiB


**Note:** current peak is caused by KAZE features allocating 8x image size utility buffers. Not sure if we can or should do anything about it: 66fb5021e9/modules/features2d/src/kaze/KAZEFeatures.cpp (L61-L68)
2024-04-17 14:19:05 +03:00

227 lines
8.0 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_invariance_utils.hpp"
#include <functional>
namespace opencv_test { namespace {
#define SHOW_DEBUG_LOG 1
// NOTE: using factory function (function<Ptr<Type>()>) instead of object instance (Ptr<Type>) as a
// test parameter, because parameters exist during whole test program run and consume a lot of memory
typedef std::function<cv::Ptr<cv::FeatureDetector>()> DetectorFactory;
typedef tuple<std::string, DetectorFactory, float, float> String_FeatureDetector_Float_Float_t;
static
void matchKeyPoints(const vector<KeyPoint>& keypoints0, const Mat& H,
const vector<KeyPoint>& keypoints1,
vector<DMatch>& matches)
{
vector<Point2f> points0;
KeyPoint::convert(keypoints0, points0);
Mat points0t;
if(H.empty())
points0t = Mat(points0);
else
perspectiveTransform(Mat(points0), points0t, H);
matches.clear();
for(int i0 = 0; i0 < static_cast<int>(keypoints0.size()); i0++)
{
int nearestPointIndex = -1;
float maxIntersectRatio = 0.f;
const float r0 = 0.5f * keypoints0[i0].size;
for(size_t i1 = 0; i1 < keypoints1.size(); i1++)
{
float r1 = 0.5f * keypoints1[i1].size;
float intersectRatio = calcIntersectRatio(points0t.at<Point2f>(i0), r0,
keypoints1[i1].pt, r1);
if(intersectRatio > maxIntersectRatio)
{
maxIntersectRatio = intersectRatio;
nearestPointIndex = static_cast<int>(i1);
}
}
matches.push_back(DMatch(i0, nearestPointIndex, maxIntersectRatio));
}
}
class DetectorInvariance : public TestWithParam<String_FeatureDetector_Float_Float_t>
{
protected:
virtual void SetUp() {
// Read test data
const std::string filename = cvtest::TS::ptr()->get_data_path() + get<0>(GetParam());
image0 = imread(filename);
ASSERT_FALSE(image0.empty()) << "couldn't read input image";
featureDetector = get<1>(GetParam())();
minKeyPointMatchesRatio = get<2>(GetParam());
minInliersRatio = get<3>(GetParam());
}
Ptr<FeatureDetector> featureDetector;
float minKeyPointMatchesRatio;
float minInliersRatio;
Mat image0;
};
typedef DetectorInvariance DetectorScaleInvariance;
typedef DetectorInvariance DetectorRotationInvariance;
TEST_P(DetectorRotationInvariance, rotation)
{
Mat image1, mask1;
const int borderSize = 16;
Mat mask0(image0.size(), CV_8UC1, Scalar(0));
mask0(Rect(borderSize, borderSize, mask0.cols - 2*borderSize, mask0.rows - 2*borderSize)).setTo(Scalar(255));
vector<KeyPoint> keypoints0;
featureDetector->detect(image0, keypoints0, mask0);
EXPECT_GE(keypoints0.size(), 15u);
const int maxAngle = 360, angleStep = 15;
for(int angle = 0; angle < maxAngle; angle += angleStep)
{
Mat H = rotateImage(image0, mask0, static_cast<float>(angle), image1, mask1);
vector<KeyPoint> keypoints1;
featureDetector->detect(image1, keypoints1, mask1);
vector<DMatch> matches;
matchKeyPoints(keypoints0, H, keypoints1, matches);
int angleInliersCount = 0;
const float minIntersectRatio = 0.5f;
int keyPointMatchesCount = 0;
for(size_t m = 0; m < matches.size(); m++)
{
if(matches[m].distance < minIntersectRatio)
continue;
keyPointMatchesCount++;
// Check does this inlier have consistent angles
const float maxAngleDiff = 15.f; // grad
float angle0 = keypoints0[matches[m].queryIdx].angle;
float angle1 = keypoints1[matches[m].trainIdx].angle;
ASSERT_FALSE(angle0 == -1 || angle1 == -1) << "Given FeatureDetector is not rotation invariant, it can not be tested here.";
ASSERT_GE(angle0, 0.f);
ASSERT_LT(angle0, 360.f);
ASSERT_GE(angle1, 0.f);
ASSERT_LT(angle1, 360.f);
float rotAngle0 = angle0 + angle;
if(rotAngle0 >= 360.f)
rotAngle0 -= 360.f;
float angleDiff = std::max(rotAngle0, angle1) - std::min(rotAngle0, angle1);
angleDiff = std::min(angleDiff, static_cast<float>(360.f - angleDiff));
ASSERT_GE(angleDiff, 0.f);
bool isAngleCorrect = angleDiff < maxAngleDiff;
if(isAngleCorrect)
angleInliersCount++;
}
float keyPointMatchesRatio = static_cast<float>(keyPointMatchesCount) / keypoints0.size();
EXPECT_GE(keyPointMatchesRatio, minKeyPointMatchesRatio) << "angle: " << angle;
if(keyPointMatchesCount)
{
float angleInliersRatio = static_cast<float>(angleInliersCount) / keyPointMatchesCount;
EXPECT_GE(angleInliersRatio, minInliersRatio) << "angle: " << angle;
}
#if SHOW_DEBUG_LOG
std::cout
<< "angle = " << angle
<< ", keypoints = " << keypoints1.size()
<< ", keyPointMatchesRatio = " << keyPointMatchesRatio
<< ", angleInliersRatio = " << (keyPointMatchesCount ? (static_cast<float>(angleInliersCount) / keyPointMatchesCount) : 0)
<< std::endl;
#endif
}
}
TEST_P(DetectorScaleInvariance, scale)
{
vector<KeyPoint> keypoints0;
featureDetector->detect(image0, keypoints0);
EXPECT_GE(keypoints0.size(), 15u);
for(int scaleIdx = 1; scaleIdx <= 3; scaleIdx++)
{
float scale = 1.f + scaleIdx * 0.5f;
Mat image1;
resize(image0, image1, Size(), 1./scale, 1./scale, INTER_LINEAR_EXACT);
vector<KeyPoint> keypoints1, osiKeypoints1; // osi - original size image
featureDetector->detect(image1, keypoints1);
EXPECT_GE(keypoints1.size(), 15u);
EXPECT_LE(keypoints1.size(), keypoints0.size()) << "Strange behavior of the detector. "
"It gives more points count in an image of the smaller size.";
scaleKeyPoints(keypoints1, osiKeypoints1, scale);
vector<DMatch> matches;
// image1 is query image (it's reduced image0)
// image0 is train image
matchKeyPoints(osiKeypoints1, Mat(), keypoints0, matches);
const float minIntersectRatio = 0.5f;
int keyPointMatchesCount = 0;
int scaleInliersCount = 0;
for(size_t m = 0; m < matches.size(); m++)
{
if(matches[m].distance < minIntersectRatio)
continue;
keyPointMatchesCount++;
// Check does this inlier have consistent sizes
const float maxSizeDiff = 0.8f;//0.9f; // grad
float size0 = keypoints0[matches[m].trainIdx].size;
float size1 = osiKeypoints1[matches[m].queryIdx].size;
ASSERT_GT(size0, 0);
ASSERT_GT(size1, 0);
if(std::min(size0, size1) > maxSizeDiff * std::max(size0, size1))
scaleInliersCount++;
}
float keyPointMatchesRatio = static_cast<float>(keyPointMatchesCount) / keypoints1.size();
EXPECT_GE(keyPointMatchesRatio, minKeyPointMatchesRatio);
if(keyPointMatchesCount)
{
float scaleInliersRatio = static_cast<float>(scaleInliersCount) / keyPointMatchesCount;
EXPECT_GE(scaleInliersRatio, minInliersRatio);
}
#if SHOW_DEBUG_LOG
std::cout
<< "scale = " << scale
<< ", keyPointMatchesRatio = " << keyPointMatchesRatio
<< ", scaleInliersRatio = " << (keyPointMatchesCount ? static_cast<float>(scaleInliersCount) / keyPointMatchesCount : 0)
<< std::endl;
#endif
}
}
#undef SHOW_DEBUG_LOG
}} // namespace
namespace std {
using namespace opencv_test;
static inline void PrintTo(const String_FeatureDetector_Float_Float_t& v, std::ostream* os)
{
*os << "(\"" << get<0>(v)
<< "\", " << get<2>(v)
<< ", " << get<3>(v)
<< ")";
}
} // namespace