sift | sift - a file and folder sifter

You are getting a exception because output argument of drawKeypoints must be 3 channels colored image, and you are initializing output to 1 channel (grayscale) image.

When using: Mat output(source.rows, source.cols); or Mat output;, the drawKeypoints function creates a new colored matrix automatically.
When using the derived template matrix class Mat_, the function drawKeypoints raises an exception!

You may replace: Mat_ output(source.rows, source.cols); with:

A simple solution may be: Iterating all keypoints and draw a "+" sign using cv2.drawMarker.

Here is a code sample:

The error comes from MLS, some values end with a question mark, "?".

You could use a relative criteria, so instead of using an asbolute value for MIN_MATCH_COUNT you can use the percentage of matching keypoints over the total number of keypoints of your model. In this way you can set a threshold based on your specific test, let's say..30% (IDK, just an example). That's what I do in a similar issue. Something like:

Build OpenCV with the following command:

After hours of profiling, I finally found out the reason: GaussianBlur.

The pipeline of SIFT algorithm is:

1. Create initial image: convert data type of source image to float, double the resolution, and do GaussianBlur (sigma=1.56)
2. Build gaussian pyramid
3. Find key points: build DoG pyramid and find scale space extrema
4. Calculate descriptors

The num of octaves is calculated according to image resolution (see here). And nOctaveLayers controls num of layers (nOctaveLayers + 3 for GaussianPyramid) in each octave.

Indeed, when nOctaveLayers increases, nums of layers and keypoints both increase. As a result, time cost of step 3 & 4 increases. However, in parallel computation, this time increment is not very remarkable (several milliseconds).

In contrast, the step 2 costs more than half of the total time. It costs 25.27 ms (in 43.49 ms) when nOctaveLayers is 3, and 51.16 ms (in 63.10 ms) when nOctaveLayers is 1. So, why is this happening?

Because the sigma for GaussianBlur() increases faster when layers are fewer, and it's critical for the time consumed by GaussianBlur(). See test below:

Feature detectors in OpenCV should be created using their static create methods. It isn't explicit in the API reference to not use the constructors, but you can see in the docs for the SIFT class that the create() methods are the only ones shown (and they show that the corresponding Python call is SIFT_create()).

The reason this is necessary is because the OpenCV Algorithm class (which Feature2D is a subclass of) is implemented using the PIMPL idiom so that SIFT is actually a base/virtual class not intended to be instantiated directly; instead SIFT::create returns a SIFT_Impl instance, as you can see in the code. The Python bindings can probably be "fixed" to mitigate this issue at some point in the future, but until then, you'll want to use SIFT_create(), BRISK_create(), and so on for all the feature detectors.

Note also that SIFT.create() and BRISK.create() and so on are valid generated names which are equivalent, but you'll find the underscore versions more common in tutorials and in people's projects, and they are currently the ones explicitly documented.

This site generates the hmac and expire url parameters value from a JS function that is using a specific algorithm. The arguments of this algorithm are depending on the epoch time which is passed as url parameter to the JS file hosting that function here. This way, the hmac value is different each time because it's processed from this file whose url is changing constantly.

This algorithm consists of bitwise and & xor like this (pseudocode):

I already did this in C++ but it is easily transferable to Python.

I am using files (matches.txt) which hold the feature point coordinates (x1, y1, x2, y2) such as trainIdx and queryIdx are similar:

x1 y1 x2 y2 (row 1 - trainIdx 0, queryIdx 0)
x1 y1 x2 y2 (row 2 - trainIdx 1, queryIdx 1)
...

In your example you would have to assing the correct trainIdx and queryIdx to the std::vector.
Then I read the file and the images and fill the overloaded std::vector with queryIdx, trainIdx and distance (which is irrelevant for drawing, so I set it to 1).

Here is the relevant piece of code in C++: