Xpano 0.15.0 - Multiblend integration

Hello everyone,

There is a new release of Xpano available, integrating the Multiblend algorithm (Multiblend):

Thanks to @prokoudine for pointing me to the Multiblend algorithm and to @sguyader and @luator for sharing example images, it was very helpful!

Compared with the OpenCV algorithm, there is better detail preservation - downscaled example pano (top OpenCV, bottom Multiblend):


100% Crop (top OpenCV, bottom Multiblend):


There is also better smoothing of uneven exposures, see Lack of detail in xpano stitched images - #5 by sguyader for an example.

On the other hand Multiblend doesn’t do seam optimization, so there can be some artifacts e.g. with moving objects (top OpenCV, bottom Multiblend, images from @sguyader):


Due to this, I haven’t made Multiblend the default yet, you can try it out in Options → Panorama stitching. I hope I’ll be able to combine the seam optimization from OpenCV with multiblend eventually.

I’d be happy to hear your feedback if you try this out.

Have a great day!

15 Likes

That’s really awesome! Great job.

1 Like

A definite improvement, excellent CONGRATULATIONS

You can compare

2 Likes

Great results, indeed now the detail is preserved, good job. Too bad regarding the lack of seam optimization, I hope you find a solve this (and please solve this before august, as I’ll return to Venice this summer :wink:)!
But for general landscapes the Multiblend algorithm is definitely a keeper.

1 Like

Haha, that is great encouragement :slight_smile: I have already tested one approach and it seems to work, so it should be ready by then.

1 Like

Cool!

I tried to build xpano on MacOS Ventura, last time it worked but this time when I launch Xpano I get the following error: Library not loaded: /opt/homebrew/opt/fmt/lib/libfmt.9.dylib

The version of fmt installed on my Mac is 10, not 9. Is there a way to make Xpano work with that version?

maybe delete the build, exiv2, install and opencv directories before rebuilding so everything is build from scratch

Indeed I had forgotten to clean my directories from previous builds.

After cleaning, I first got a fail with the exiv2 that is cloned and built from git, but it worked building Xpano against the already installed exiv2 from brewer.

In fact I don’t find the multiblend stitching option… Did I miss something?

Congrats on the build! :slight_smile:

Sadly, the multiblend algorithm only works on x86 currently, as it uses SSE instructions. I’ll update the xpano readme about this (GitHub - krupkat/multiblend: Fork of multiblend: http://horman.net/multiblend/). I want to have an arm port implemented next.

1 Like

fantastic work, thanks a lot!!

1 Like

Xpano 0.15.0 Works fine on linux Manjaro, the multiblend algorithm: is definitely a plus regarding texture over opencv.
It would be cool to know why the original opencv algo smears the detail so much, maybe it could be fixed?

That would be fantastic.

I tried to rule out user error on my side by trying out the official opencv example (OpenCV: samples/cpp/stitching_detailed.cpp) with various parameters and I can’t reach the results given by Multiblend.

Both algorithms are based on the same principles (http://persci.mit.edu/pub_pdfs/spline83.pdf), but there is a lot of details where it can go wrong.

I looked through bugs on opencv github and didn’t find any related issues, so I’ll probably create one when I have time.

2 Likes

I released a version with Multiblend + OpenCV seam optimization:

The results seem quite nice, I’ll probably make this the default unless there are any further issues with it.

7 Likes

I just tried on Linux, and I think it’s perfect! Great job Tomas!

1 Like

Awesome!

1 Like

Thanks a lot everyone for the nice comments!

For those interested, there is now support for Multiblend on ARM (and others possibly):

I have tested on an ARM chromebook and it worked well, it should be working on those M1+ chips now. Enjoy!

3 Likes

Great software, thank you @krupkat !
I think stitching is at least as good as with ICE

1 Like

Works fine on my Mac mini M1, great job again!

1 Like