Page 1 of 1


PostPosted: Wed Mar 23, 2011 8:46 am
by wjh31
Ive had a look at a couple of the posts around on neutral hazer, and i cant quite see why this is better than e.g playing with levels in the editing screen or in post processing. Haze tends to be worst right on the horizon, but the posts seem to apply that the same level of neutralhazing is applied to all images.

Anyone able to provide a little more insight on my misconceptions?

PostPosted: Wed Mar 23, 2011 8:53 am
by AlexandreJ
The difference between neutralhazer and level editing is the quality of the mask used for that.
Just use the ALT key while beeing in the haze plugin to learn about the mask : we calculate a pixel accurate depth mask to be able to remove haze. When doing that using PS levels, you would never spend time to make such an accurate mask.

PostPosted: Wed Mar 23, 2011 6:48 pm
by wjh31
I think i understand. So you tell it how much to increase the contrast in one area, and the plugin tries to figure out how far away each pixel is in the other images and applies the appropriate level of the effect?

PostPosted: Thu Mar 24, 2011 8:22 am
by Castillonis
Yes, I believe that Alexandre uses a method that was developed by Raanan Fattal which is described in his paper "Single Image Dehazing"

This method depends upon the shadows not being correlated with the haze attenuation and airlight. It uses this assumption to create a depth map. This method is much more practical because only one image is needed as well as no special equipment such as circular polarizer filters or depth maps as input. The radiance values are decreased proportional to the distance because some of the light is scattering when it hits particles in the atmosphere. The airlight is light that has been reflected off of other particulate and is adding to the radiance value.

A really interesting paper was authored by Neel Joshi and Michael F. Cohen of Microsoft research that uses multiple images. They employ algorithms such as Lucky imaging from astronomy.

I just worked on a job where I was very challenged by haze. I had to capture the image at about 1pm because of the necessary position of the sun in a very wide horizontal field of view. Unfortunately many of the subjects in my panorama were very far and extremely effected by heat shimmer and haze. I had to reduce the resolution by aproximately 50% to ensure that the image would retain reasonable quality at 100% zoom. I believe that the only method that would help me produce a better image would involve tools similar to what Joshi and Cohen developed. Hopefully this method my work with fewer images than they used in their research.

PostPosted: Thu Mar 24, 2011 8:44 am
by wjh31
cool, if the output looks good i might have to go back and re-render a whole bunch of my images

PostPosted: Thu Mar 24, 2011 10:20 am
by AlexandreJ
Yes, Castillonis, we used this paper as inspiration. THe general idea is the same, but we changed everything after that to be able to reach real-time preview.

PostPosted: Fri Mar 25, 2011 3:42 am
by Castillonis
Alexandre and Lionel, I want to thank both of you for working arduously to improve the Autopano stitcher and especially the dehaze filter which just helped me with my last job :)

PostPosted: Fri Mar 25, 2011 9:08 am
by AlexandreJ
Thanks. As I always say, just don't hesitate to give us any feedback about something that is hard to do currently and we'll find the solution. The history of haze development is in fact that : I had some issue with some haze in my panorama, I did google for a solution, found the paper, coded it and done :)