OpenCV: Replace ilLoadImage with cv::imread

Question

I'm using hessgpu for efficiently compute hessian-affine SIFT descriptors using GPUs.

In this project, Devil is used for reading images: if SiftGPU::RunSIFT(const char *imgpath) is called (implemented here), then ilLoadImage is used in GLTexImage.cpp for reading images (as RGB images).

However, in my project I use cv::imread to read images. The project provides SiftGPU::RunSIFT(int width, int height, const void * data, unsigned int gl_format, unsigned int gl_type) to compute descriptors from data provided directly from the user.

So I tried:

cv::Mat img = cv::imread("image.jpg", cv::IMREAD_GRAYSCALE);
sift.RunSIFT(img.cols, img.rows, img.data, GL_LUMINANCE, GL_UNSIGNED_BYTE);

But this produce slightly less keypoints than sift.RunSIFT("image.jpg");. I tried to use:

cv::Mat img = cv::imread("image.jpg");
sift.RunSIFT(img.cols, img.rows, img.data, GL_RGB, GL_UNSIGNED_BYTE);

But this produces 0 keypoints, so something very wrong happens. I think:

1. iLoadImage uses RGB image, while the only working method that I found up to now using cv::imread works only with grayscale images.
2. It's possible that devil use a different process to read images than OpenCV, espcecially for RGB images (since using the second approach produced 0 keypoints).

How can I do the equivalent of ilLoadimage using cv::imread?

Answer 1

Changing the runSift line to:

sift.RunSIFT(img.cols, img.rows, img.data, GL_BGR, GL_UNSIGNED_BYTE);

Solved the problem and produced the exact amount of descriptors.

Answer 2

It seems that something's wrong with types, try this one:

Mat img = imread("image.jpg"); 
sift.RunSIFT(img.cols, img.rows, img.data, GL_BGR, GL_3_BYTES); 

Or this one

cv::Mat img = cv::imread("image.jpg", cv::IMREAD_GRAYSCALE);
sift.RunSIFT(img.cols, img.rows, img.data, GL_RED, GL_UNSIGNED_BYTE);