Higher Resolution Images

Skip to first unread message


Nov 27, 2018, 6:41:39 PM11/27/18
to Theia Vision Library
I've noticed that the default SIFT parameters with theia using the DENSE multiplier/divider becomes more "SPARSE" as image resolution is increased (which i believe is expected). (More here, http://www.vlfeat.org/api/sift.html and here, http://www.vlfeat.org/overview/sift.html)

For example, the average distance between detected key points appears to increase (qualitative observation) with photos at resolution of 7360 x 4912, vice my typical 1920x1080 or 4K+ datasets.

I am curious as to what peak/edge thresholds are used by others in this group when going beyond 1080p and 4K? How about reducing the first octave index (to -2) or increasing the number of levels per octave (to 4?). I'm seeing more key points by doing so, but am now waiting around results from matching.

Any thoughts around this would be helpful since there are a lot of pixels in my input imagery that are not detected, consider only ~5K features per image @ 7360x4912. These are pretty simple aerial images shot in nadir with ~80% overlap and plenty of unique features.

Thanks in advance,


Charles O

Nov 28, 2018, 9:18:20 AM11/28/18
to Theia Vision Library
It looks like appending two vectors, one consisting of Akaze Sparse matches and one with Sift Dense matches before reconstruction yields more verified 3d points in areas of interest. As Akaze picks keypoints SIFT doesnt.  This sort of achieves the desired effect.  With modified Sift parameters i am getting around 12K features per high res. image now, which is about 30% more than the defaults and the reconstruction still looks good.

You received this message because you are subscribed to a topic in the Google Groups "Theia Vision Library" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/theia-vision-library/J4blQZ5PJaQ/unsubscribe.
To unsubscribe from this group and all its topics, send an email to theia-vision-lib...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
Reply all
Reply to author
0 new messages