AGAST detector is ~30x slower than FAST detector

main.cpp

// Output (DETECTOR, num detected points, total milliseconds)
// AGAST, 3715, 31672
// FAST, 3735, 1025

int main(int argc, char *argv[])
{
	cv::Ptr<cv::AgastFeatureDetector> agast =
		cv::AgastFeatureDetector::create(10, true, cv::AgastFeatureDetector::OAST_9_16);

	cv::Ptr<cv::FastFeatureDetector> fast =
		cv::FastFeatureDetector::create(10, true, cv::FastFeatureDetector::TYPE_9_16);

	cv::Mat img = cv::imread("example.jpg", cv::IMREAD_GRAYSCALE); // 512x512
	size_t NumLoops = 500;
	double startTicks, elapsedTicks, elapsedMilliseconds;

	size_t agastPoints = 0;
	startTicks = (double) cv::getTickCount();
	for(size_t nn = 0; nn < NumLoops; nn++) {
		std::vector<cv::KeyPoint> keypoints;
		agast->detect(img, keypoints);
		agastPoints += keypoints.size();
	}
	elapsedTicks = (double) cv::getTickCount() - startTicks;
	elapsedMilliseconds = 1.0e3 / cv::getTickFrequency() * elapsedTicks;
	std::cout << "AGAST, " << (int)((double)agastPoints/NumLoops) << ", " << (int)elapsedMilliseconds << std::endl;

	size_t fastPoints = 0;
	startTicks = (double) cv::getTickCount();
	for(size_t nn = 0; nn < NumLoops; nn++) {
		std::vector<cv::KeyPoint> keypoints;
		fast->detect(img, keypoints);
		fastPoints += keypoints.size();
	}
	elapsedTicks = (double) cv::getTickCount() - startTicks;
	elapsedMilliseconds = 1.0e3 / cv::getTickFrequency() * elapsedTicks;
	std::cout << "FAST, " << (int)((double)fastPoints/NumLoops) << ", " << (int)elapsedMilliseconds << std::endl;
}

(Using OpenCV 3.0)

Illustrates that for a 500 loop run on an example image, the AGAST detector took a total of 31.7 seconds while the FAST detector took 1.03 seconds. Number of points detected by each is approximately the same (3715 for AGAST and 3735 for FAST).

Submitted as Issue #5278.