Mårten Björkman

I think I found the problem. First you need to read the image in grayscale. Add 0 as an argument to imread, i.e. cv::imread(strPath, 0). The code doesn't recognise that...

Since memory allocation is so costly, I added a function called AllocSiftTempMemory that preallocates memory. If you have a real-time system with images coming from a video camera, you might...

I think it depends on the pitch of allocated images. I believe that the number of pixels in width has to be divisible by 128, so during allocation on the...

Some years back, I managed to make it run on a Jetson TK1. It was surprisingly easy, but there is no guarantee that it works as easily with the current...

Yes, there is a bit of cleaning up to do there. Sometimes when I detect oddities in the output, I add an unnecessary synchronization just in case. In fact those...

Yes, that is a correct observation. In the earliest version of the code, I used the layer corresponding to the scale for orientations and descriptors, but in 2010 I changed...

While (not) preparing for a lecture this morning, I had to do some testing and went back to the pre-2010 version of the code. I did the same observations you...

That is weird. Could you upload the image? I cannot verify it. It might be something related to the unusual image size, but that seems like something I should have...

By the way, MatchSiftData uses both siftData1 and siftData2. It will create per feature cross-references, which means that FindHomography and ImproveHomography won't need siftData2.

Unfortunately, I cannot verify it on my machine with the images from the data/ directory, even with your showCorrespondence function. Weird.