Updating STAR detector and FREAK descriptor to work with large and/or 16-bit images

[seth-planet]

A project that I've been working with involves large, 16-bit, images. The current implementations of the STAR feature detector and the FREAK feature descriptor don't support either large or 16-bit images. I've written a set of changes to support both. Much of the work goes into adding more bits to the integral image when appropriate. I'll discuss what needed doing:

• In STAR, the method of detecting if the matrix was color didn't make sense. I'm using channels() instead of type() now.
• In STAR, he integral image data type needed to be changed for two reasons. If each cell can have the value of 256, the integer type could overflow when the image is greater than 8 megapixels. If you are working with 16-bit imagery imagery instead of 8-bit, this will happen at much smaller image sizes. So, I've used templates to use a 32-bit integer under the normal case, and 64-bit double for large images and/or higher bit depth.
• In STAR, SSE2 is only enabled if the data type is the original 32-bit integer. It was my aim to support the new data types (shorts & doubles) without removing the existing SSE usage.
• In integral(), in sumpixels.cpp, I added templated support for working with 'short' & 'ushort' precision input. As mentioned before, it's appropriate to use double precision when working with 16-bit images (so that's what I did).
• In FREAK, I broke most of computeImpl() into a new function, computeDescriptors(). I did this so I could use templates and thus keep all the data type permutations sane.
• In FREAK, SSE2 is only used if the average value is expected to be an 8-bit number. This is the original usage, and it should be just as fast as before. 16-bit data falls back to the scalar code.
• In FREAK, meanIntensity() now uses templates.
• In FREAK, the return value of meanIntensity() is 'int' instead of 'uchar' for compatibility between possible data types. It should make no speed difference.

[SpecLad]

There are a lot of commits here that don't belong.

[seth-planet]

How do I get the branch synced with the current master? I'm a git noob, I'm not sure what attempt number we're on now, and this was the best I could come up with.

Original bugfix:
http://code.opencv.org/issues/3322

On Oct 21, 2013, at 7:30 AM, Роман Донченко notifications@github.com wrote:

There are a lot of commits here that don't belong.

—
Reply to this email directly or view it on GitHub.

[apavlenko]

The windows build fails due to level 4 warnings in freak.cpp

[seth-planet]

Hm. How can I get this only on Windows?? I have all of the compile warnings & errors fixed.
"
Features2d_DescriptorExtractor_FREAK.regression (108 ms) :
Average time of computing one descriptor = 3.09155e-007 ms.
Max distance between valid and computed descriptors 348>12 - bad accuracy!
"

[SpecLad]

@seth-planet Looks like you submitted the same branch for this PR as for the old one (#1639). You need to rebase this branch locally on top of master, and then push -f it to your fork.

[apavlenko]

now it fails FREAK test (http://build.opencv.org/builders/precommit_windows/builds/6528)

[seth-planet]

Can I run the FREAK test on Mac (or Linux)? For my own test cases, I'm getting identical results between master and my branch. I'd rather not have to install Windows just to replicate this bug.

[seth-planet]

@SpecLad So, I tried to follow your rebasing instructions. Hopefully it helped things.

[SpecLad]

You have even more commits now, not less. Try interactive rebase or perhaps the --onto option. You'll know you got it if git log --oneline origin/master..lg_16_bit_img_features only shows commits you made.

Can I run the FREAK test on Mac (or Linux)?

Not sure what you mean here. Every test can be run on any platform.

[seth-planet]

When i try to run the tests on my Mac, they all fail due to missing an image. I can't find any documentation on this.

But the test reportedly only fails on windows anyway.

Sent from my iPhone

On Oct 23, 2013, at 1:13 AM, Роман Донченко notifications@github.com wrote:

You have even more commits now, not less. Try interactive rebase or perhaps the --onto option. You'll know you got it if git log --oneline origin/master..lg_16_bit_img_features only shows commits you made.

Can I run the FREAK test on Mac (or Linux)?

Not sure what you mean here. Every test can be run on any platform.

—
Reply to this email directly or view it on GitHub.

[SpecLad]

Sounds like you're missing test data. You need to set the OPENCV_TEST_DATA_PATH environment variable to point to the testdata directory inside the opencv_extra repository.

[seth-planet]

Ok, I worked my way through all the code, and figured out what the 'bug' is:

    if( image.channels() > 1 )
        cvtColor( image, grayImage, COLOR_BGR2GRAY );

I'm automatically changing color images to grayscale. If I comment that out, the test passes. However, I would think that conversion to grayscale would be preferred. (That's how it's done in the STAR detector.) What is the preferred behavior?

[SpecLad]

@vpisarev I believe that's a question for you.
@seth-planet You still need to get rid of all those unrelated commits, by the way.

[seth-planet]

I'll wait for @vpisarev and create a clean pull request.

[seth-planet]

hello?

[apavlenko]

@vpisarev, could you please look at it?

[seth-planet]

ping

[SpecLad]

@seth-planet Actually, you might want to get rid of excess commits before the review - otherwise the diff is cluttered and it's not clear at all which changes are yours. Plus, a green build instills more confidence in your code. 😉

[seth-planet]

Clean pull request:
#1932