[ptx] autopano: first attempt at semi-automatic panorama stitching

[Ed Wildgoose]

Yili Zhao wrote:

>ED wrote:
>  
>
>>but do you think that it will work well for 8mm fisheye lens,
>>or even just very wide angle lenses?
>>    
>>
>maybe for 8mm full circular lens, if we remap it to equirectangular
>format, constructing the spherical panorama becomes a pure translation
>problem.You can see this process from: 
>http://www.path.unimelb.edu.au/~bernardk/tutorials/360/background/process.html
>Thus phase correlation and klt can be used. Other wide angle
>images also need to be remapped before correlation search.
>
>  
>
You would also perhaps have problems with handheld shots if the rotation 
was more than a few degrees off center.  I have seen a few papers on 
rotational invariance, I wonder if you can make phase correlation 
rotation invariant..?  I suspect not?

>I think phase correlation can be only used to constructing single row panos.
>For multiply row panos, three parameters (yaw, pitch and roll) method should be used.  
>  
>

Sorry for being ignorant, but why not?  Would you not simply remap all 
the images assuming some fixed overlap top/bottom and then phase 
correlate, then switch top and bottom, remap appropriately and re-run 
the phase correlation - pick the best answer?


To be honest, if it was only the initial overlap which was the problem 
then I suspect that most users would be ecstatic if they merely had to 
visually drag the pictures into rough alignment and let the auto 
algorithm do the rest!  PTGui has a nice mode which lets you do this (as 
an example of an implementation).

Pablo, how accurate does the inial alignment have to be for your 
optimisation approach to work?  How good is it at rejecting invalid 
matches?  For example is a very simple algorithm which merely assumes a 
30 percent overlap on each picture, uses no phase correlation and just 
tries all permutations of images until something matches, going to work?

Seems really exciting anyway.  Thanks all