[ptx] Questions on Lowe's SIFT algorithm

[Xianyong Fang]

Sebastian,
 
    I have uploaded the test images in the public ftp server of our laboratory. The ftp address is 210.32.131.214. Please login in anonymous way and check the directory:\incoming\SIFT_TestImages\. I have uploaded four image pairs. If you can not download them, please let me know and I will sent them to you by email.

    As you said, for panorama images, it is okay. Maybe Lowe purposely omitts some details. Or maybe there indeed exist some shortcomings for Lowe's SIFT algorithm. As far as I know, there are really few algorithms can be always strong on image processing. So maybe interaction is needed in the finally implementation.

    To PCA-SIFT, please check this address:http://www-2.cs.cmu.edu/~yke/pcasift/. It is quite new but according to my implementation (actually I just inserted the pca-implementaion into my originally program) it is hard to say that it is more robust than Lowe's, esp. to rotated and scaled images.

	To my code, most of it is copied from your implementation,;). The only modification I applied is from PCA-SIFT which takes the scale shift into consideration when doing low contrast threasholding. But this improves the quality little according to my observation.

    You really do a wonderful job. I will test your implementation when you upload it. But actually I am afraid that you need us to try hugin after running your code. I feel it is not convienent and I am not familar with hugin,*-*.  
 
Xianyong
 
===================== Original Message =======================
From: Sebastian Nowozin
To: Xianyong Fang; ptx at email-lists.org
Sent: 2004-05-16 00:08:24
Subject: Re: [ptx] Questions on Lowe's SIFT algorithm
 
    
Hi Fang Xianyong,


On Sat, May 15, 2004 at 10:24:20PM +0800, Xianyong Fang wrote:

> I find that the SIFT algorithm of Lowe's paper is not robust to scale
> transform. When I try to register two images with different scales, there are
> few (or not so many correct) matches existing. To images with rotation and
> different scales, it works even worse.  But when I apply the sift algorithm
> to images with simple shift or small rotation, it works fine and almost all
> the returned matches are correct.  Does anyone have the same experience with
> me?

I have not tested in detail, but only some small tests with scale-different
pictures. It always worked fine. Can you upload a testcase somewhere? And,
which implementation do you use?

If it really is not scale-invariant as much as it should be, I still think for
panorama images it is no problem at all.


> I also find that in Lowe's IJCV paper "Distinctive Image Features from Scale
> Invariant Features" the testing images showed almost have the same scale. I
> can't help thinking that maybe his algorithm is hard to copy with images of
> different scales.

While the paper is good in general and I used it exclusively for my SIFT
implementation it has some shortcomings for practical implementation purposes
(some small but important steps must have been ommited in the paper, though I
do not know why or if my conclusions are wrong at some points).


> In addition, I also try to intergrate "PCA-SIFT: A More Distinctive
> Representation for Local Image Descriptors" into the matching process and
> find the same problem. 
 
Sounds interesting, URL to that paper?


> Xianyong

ciao,
Sebastian

PS: I finished adding RANSAC and some image model to my autopano-sift toolbox,
and would like to have your and the other people's opinion on the image align
model I use. I will upload a small sketch and the new version tomorrow or on
Monday and post to the list then.

-- 
nowozin at cs.tu-berlin.de --- http://user.cs.tu-berlin.de/~nowozin/


==============================================================