SURF_CUDA produces nan values

edit

I'm using SURF_CUDA to describe around 5k images from the Oxford dataset. All the images are resized keeping the aspect ratio such that the largest size is a given amount of pixels (usually I set 400). This is the code that I'm using to compute descriptors for each image:

void SURF_CUDAOpenCV::ComputeDescriptors(cv::Mat &img, cv::Mat1f &descriptors){
    cv::cuda::GpuMat imgGpu, descriptorsGpu;
    std::vector<float> vf_descriptors;
    std::vector<cv::KeyPoint> keypoints;
    imgGpu.upload(img);
    if(imgGpu.empty())
        throw std::runtime_error("Error uploading gpuImg");
    surf_cuda(imgGpu, cv::cuda::GpuMat(), keypoints, vf_descriptors);
    descriptors = cv::Mat1f(vf_descriptors);
    descriptors = descriptors.reshape(1,(int)vf_descriptors.size()/surf_cuda.descriptorSize());
    for(size_t i=0; i<vf_descriptors.size(); i++)
        assert(!std::isnan(vf_descriptors[i]));
}

For some reason, the assert fails for some images, which means that some descriptors values are nan, which is obviously an error.

I noticed that resizing the images to 500 delay the error (the first image containing at least a nan is the 1902-th, instead of 159-th), however I really wonder why this happens.

The error came out because I'm using these descriptors for k-means and some centroids values are nan.

UPDATE:

I found out that not resizing the images the error doesn't happen and everything is fine. The function that I use to down-resize the image keeping the aspect ratio is this one:

void resize(cv::Mat &img, float targetSize){
    float max = std::max(img.rows,img.cols);
    if(max>targetSize){
        float ratio = targetSize/max;
        cv::resize(img,img, cv::Size((int)(img.cols*ratio),(int)(img.rows*ratio)),0,0,cv::INTER_CUBIC);
    }
}

Why this happens?