Mårten Björkman

I made some tests with 17-point filters for the detector, rather than the 9-point filters that I've used so far. It's true that with a 9-point filter you truncate a...

The code is based on a CPU version I wrote about 15 years ago, but quite a bit has changed since then when it comes to the implementation, so the...

The reason is to make it as scale-invariant as possible. From the theory, the gradients for the descriptor should be computed at a scale that corresponds to the scale at...

To me, it doesn't seem like a good idea to use the original resolution for the descriptor. Large features, detected at a coarse scale, will then contain high-frequency components that...

Do you have an image illustrating the problem? SIFT features are often found in the weirdest of places.

I fail to see what the problem might be. It indeed looks like as if the coordinates are scaled up by a factor of two.

Is it just the order of features that is different or is it the features themselves? Since features are extracted in parallel, you never know in which order they will...

SiftPoint.score is only used for feature matching. If you match features and it changes like that, then something is definitely wrong. If you don't match, then you should instead use...

It's done in InitSiftData. The reason for not doing that as part of ExtractSift is because cudaMalloc is so terribly costly. It easily takes more time than the extraction itself....

I will take a look at it. I have one direct comment. Lowe uses an ambiguity measure for matching to remove features that have more than one good match in...