search

Ashwinning / libmv

Automatically exported from code.google.com/p/libmv
MIT License
0 stars 0 forks source link

SURF is unreliable #11

Open GoogleCodeExporter opened 9 years ago

GoogleCodeExporter commented 9 years ago 
What steps will reproduce the problem?
1. Download images from Issue 1
2. Use TVR to compute SURF points
3. Note that the SURF points are sometimes in homogeneous areas, and that
there are near-duplicates.

What is the expected output? What do you see instead?

SURF should do a better job of finding distinct areas, and should not find
keypoints in homogeneous areas. Also, multiple close keypoints should be
pruned.

Please use labels and text to provide additional information.

Original issue reported on code.google.com by mie...@gmail.com on 1 Jul 2009 at 10:16

GoogleCodeExporter commented 9 years ago 
Actually when testing Bundler I've switched it from using SIFT to SURF and the 
results 
were awful too.

Original comment by bomb...@gmail.com on 1 Jul 2009 at 10:28

GoogleCodeExporter commented 9 years ago 
Note that with the change to integers for the integral image, since the pixel 
values 
are scaled by 255, so are the filter results. The blob response is scaled by 
255*255 
(because the determinant has units of integral image^2). Thus, the threshold 
for 
whether or not something is a maxima probably needs to change.

Original comment by mie...@gmail.com on 1 Jul 2009 at 5:41

GoogleCodeExporter commented 9 years ago 
I think that images were already scaled by 255.  We were using floating point 
images
but they values where copied directly from a QImage.  So they where between 0 
and 255.

By the way, Issue 1 seems to be fixed even with floating point images.  If I 
now set
the type of both the image and the integral_image inside surf to Matf, then 
BoxFilter
no longer gives negative values.

Original comment by pau.garg...@gmail.com on 3 Jul 2009 at 2:27