Ask Your Question

Removing outliers from goodFeaturesToTrack using the x84 method

asked 2013-01-10 10:15:15 -0600

gary gravatar image

Hi, I found this paper that describes what looks to be a pretty great outlier removal method. Unfortunately the formulas are a bit over my head.

Does know if this x84 method is implemented in OpenCV? I have some occluded keypoints due to "foreground" motion (similar to how the road sign is occluded by the teddy bear in that PDF) and I need to discard them.

Any help would be much appreciated!

edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted

answered 2013-01-10 12:39:02 -0600

Adi gravatar image

Instead of using all the feature points, reject all points that are more than k Median Absolute Deviations (MADs) away from the overall median.
They suggest a value of k=5.2, under the hypothesis of Gaussian distribution, as it corresponds to about 3.5 standard deviations, and contains more than the 99.9% of a Gaussian distribution.

edit flag offensive delete link more



Thanks so much for the info, and nice job on the OpenCV logo! That sounds like it makes sense, but unfortunately I have no idea where to even start on the OpenCV side of things. I was hoping for some kind of X_84 flag/parameter in the TranslationBasedLocalOutlierRejector class. Say I have an array of images, along with corresponding prevPoints and nextPoints from calcOpticalFlowPyrLK. How would I go about computing the MADs and the gaussian distribution, etc. and in turn removing these outliers? Thanks again.

gary gravatar imagegary ( 2013-01-10 13:07:07 -0600 )edit

I don't know the TranslationBasedLocalOutlierRejector class, but apparently it does not support the X84 algorithm and you'll have to implement it yourself. I suggest your read the paper carefully, the details are explained within it.

Adi gravatar imageAdi ( 2013-01-13 01:50:37 -0600 )edit

Alright, thanks.

gary gravatar imagegary ( 2013-01-13 04:07:17 -0600 )edit

Question Tools

1 follower


Asked: 2013-01-10 10:15:15 -0600

Seen: 1,389 times

Last updated: Jan 10 '13