²RE : [ptx] autopano1.03 comparison to autopano-sift1.4

[alexandre jenny]

Hi Ian,

First thanks you for this comparison. It's interesting to see how
the both system are working.

I used this testing case as a startup for the improvement of my
system. And I achieved far better results as before.

If you have time to check again the new beta release of autopano_v1.03beta2,
it could be a good measure of the new algorithm inside.

Here it is :
  http://www.le-geo.com/kolor/autopano/autopano_v103beta2.zip

I suggest you to run it directly with default parameters
(but with /keys:100 for getting 100 keys).

And then use /size:2000 /keys:100 also. Just leave the other switch.

Regards,
  Alexandre Jenny


> I've run a small comparison of control point creation for 2 images 
> 2048x1536 taken from a moving cable-car and asking for 100 control 
> points. A difficult task anyway, but the comparison was very clear.
> 
> Autopano-sift ran *much* slower than autopano1.03 (or 1.02)- 
> I did not 
> use a stop watch, but about 3 times as long as autopano.
> 
> Autopano-sift at size:2000 found me 100 control points and when 
> optimised in Hugin the average control point distance was 
> 0.74 pixels, 
> maximum 1.67.
> 
> Autopano1.03 at size:2000 found 44 control points and when optimised 
> in Hugin the average control point distance was 33 pixels, maximum 53 
> pixels.
> 
> Autopano-sift at size:800 found 64 control points and when optimised 
> in Hugin the average control point distance was 0.75 pixels, 
> maximum 2.14.
> 
> Autopano1.03 at size:800 found 29 control points and when 
> optimised in 
> Hugin the average control point distance was 14 pixels, maximum 25 
> pixels. The optimisation gave a warped image and on closer inspection 
> I found 5 incorrect control point matches. Removing them gave 
> a better 
> image, but did not significantly improve the control point error.
> 
> Summary:
> For this test the autopano-sift algorithm generated more 
> control point 
> matches, and only found "proper" match points.
> Autopano1.03 was much much faster, but the output file still required 
> lots of manual changes to get a good image.
> 
> Is this comparison similar to anyone else's findings?
> 
> 
> regards,
> 
> Ian Sydenham
>