MitsunagaNayarCRF is complete? about hdr_toolbox HOT 34 CLOSED

you are correct, I am working on this part.

from hdr_toolbox.

ok,thanks,I also test this part

from hdr_toolbox.

I made some changes, especially I added more examples for building

from hdr_toolbox.

ok, i suggest you can add the Debevec's dataset
http://www.pauldebevec.com/Research/HDR/SourceImages/Memorial_SourceImages.zip

from hdr_toolbox.

I have tested this dataset
with new version code of demo_build_hdr
but will ouput very strange result....

from hdr_toolbox.

This can not be done, because permissions are required and needed to be granted.
I will add more example soon.

You need to manually add exposure values, there are not in the images' metadata.

from hdr_toolbox.

yes, i have manually add exposure values, but seems something still wrong
but i think its ok, its public available on its own website
http://www.pauldebevec.com/Research/HDR/
maybe only need to cite?

from hdr_toolbox.

No you need explicit permission. I will add another dataset with many exposures. I will check that dataset meanwhile.

from hdr_toolbox.

The issue with the dataset is the blue parts in the stack. I guess images were aligned using Homography. The blue parts create issue when computing the CRF. A solution to this issue is to crop the stack to avoid to include those blue parts in the estimation of the CRF.

from hdr_toolbox.

I have tried many weight function
but seems all don't work
the image after tonemapping looks like all black
I find the wired warning of rank deficient of 1155,1155,1148 appears
and I really don't know why when i choose nSamples as 900, there will still rank deficient,
it seems not uniform to the inequality Debevec states in his paper.
I have the cropped version, but in the office, maybe i can give it a try tomorrow

from hdr_toolbox.

If you check the CRFs, it is completely wrong for the blue channel. This is due to the fact that the blue parts are included in the CRF estimation.

from hdr_toolbox.

The CRF i get all wrong
looks like this.....

from hdr_toolbox.

ah, you are using the wrong Exposure time values, please try 1/exposure_time as stated in the .txt file

from hdr_toolbox.

Sorry for my stupid mistake....
I will try cropped version in the office
thanks

from hdr_toolbox.

it can happen, you are welcome

from hdr_toolbox.

For more robust, stated in the literature
should correlated the color channel relationship before and after HDR construction(I mean in the generation of the radiance map)
How about the rank deficient problem?

from hdr_toolbox.

More robust methods will be added in the future and they are planned in the road map. However, it takes time to add and test them.

from hdr_toolbox.

Thanks, it's a good toolbox!!!
Maybe the rank deficient cause from the sampling?

from hdr_toolbox.

even I use the dataset cropped
the result looks like the similar wrong results as the uncropped ones
I also try many various types of weighting funtion

from hdr_toolbox.

I tried the built-in tonemap function in matlab
test result show that still show error in building the radiance map or construct the CRF

from hdr_toolbox.

I tried to build this dataset with another program, PictureNaut 3.2, but it fails in the middle of the process.

from hdr_toolbox.

Sorry, but i mean use the cropped version, not original one
i have checked another HDR script on github, it works well
https://github.com/cloudbopper/vision-hdr

Since the CRF reconstructed also used Debevec's method, I think there is a bug in this part
But the tonemapping methods is great, I use that to tonemap the hdr I got through above utl's script

Also,I found the cropped version of this on OpenCV extra, for your to test
https://github.com/Itseez/opencv_extra/tree/master/testdata/cv/hdr/exposures

from hdr_toolbox.

To be honest with you, I tested that sequence with the cropping and it works. The CRF is well recovered with cropping; I used original Debevec stack with cropping the stack at x_coord=[28, 466] and y_coord=[58, 710]; just to be sure to not have the blue bit in it.

I noticed that that noise can appear if the sequence is built in the linear domain, try to use BuildHDR using 'log' instead of 'linear'; which is more robust especially in cases with many images at different exposures in the stack.

I will update the example soon.

from hdr_toolbox.

Thanks, It works now, but another observation using the same smooth term
suggest that maybe some error occures due to the truncation, i try to find it now
Because when you carefully observe the tonemapped result, still some green noise spread

from hdr_toolbox.

which truncation are you referring to?

from hdr_toolbox.

PS: I will modify from now on the develop branch, so please have a look to it.

from hdr_toolbox.

Sorry, but I am wrong
the bugs may comes from the sampling
I replace the sampling process
then the rank deficient it not occur again
and the output looks like without green spot liked noise
even if i used a bad sampling method

from hdr_toolbox.

in computeCRF file

I place following code inside the for loop for construct the CRF
and replace the function call of gsolve
N=256

from hdr_toolbox.

The Grossberg and Nayar sampling may generate rank deficiency, but the output is more stable with few samples. To get the same with spatial sampling (which was already implemented in the toolbox, and I now renamed as SpatialSampling) does not have the rank deficiency issue but it requires many samples to have stable results (even though it can happen with regular sampling).

When I do the cropping (for avoiding the blue parts for recovering the CRF), I simply do not get the greenish thingie. I will add the option of the spatial sampling and that is all.

Note that I am doing changes at the moment in the develop branch. Please check that for future.

from hdr_toolbox.

ok,thanks for you explanation
I will check the script again

from hdr_toolbox.

what i mean a little green noise
But it should not cause rank deficient,
or need to avoid in advance?
More robust method should have lower quality result
Weird, right?

from hdr_toolbox.

Ah ok, I got now what you meant, I analyzed closely the image and I spotted the subtle greenish noise. I will investigate it now.

I will analyze the situation with synthetic data. Thanks again for your devoted help in this.

from hdr_toolbox.

No thanks,
It's an awesome project
I'm glad to involve in it

from hdr_toolbox.

you are welcome.

from hdr_toolbox.