This is the official course blog for 16824 Learning-Based Methods in Vision (Spring 2010). Class is taught by Martial Hebert and held on Mondays and Wednesdays 9:00am-10:20am in GHC 4211.
I really like this paper for going into the dirty details, especially since there so many parameters that no one likes to practically deal with. I think more work needs to be done like this and to revisit older models with better representations/features. One such recent paper that is burned into my mind is Lubor Ladicky & others, Associative Hierarchical CRFs for Object Class Image Segmentation, ICCV 2009 http://research.microsoft.com/en-us/um/people/pkohli/papers/lrkt_iccv09.pdf where (while not the point of the paper) by adding just HOG features to a standard CRF model greatly improves over 4 other recent works.
Anyway, the conclusion to me is that by letting the classifier train over lots of random data, you remove any training bias from your trainingset that your interest point detector might focus on that may not be present in the testset and essentially leave it to the learning algorithm figure out the meaningful and useless features. I think one way to further solidify this hypothesis is to show another axis of classification performance vs object % of the image. Judging from the couple images shown Fig 2, I can buy that a random sampling performs well when the object occupies > 40% of the image, but when it is very small, like for the people in 2nd & 4th image, i am wondering if the interest point detectors work better.
Another argument for random quantization: projecting your data on random dimensions finds the underlying distribution http://cseweb.ucsd.edu/~naverma/RPTrees/publication.html
I really like this paper for going into the dirty details, especially since there so many parameters that no one likes to practically deal with. I think more work needs to be done like this and to revisit older models with better representations/features. One such recent paper that is burned into my mind is
ReplyDeleteLubor Ladicky & others,
Associative Hierarchical CRFs for Object Class Image Segmentation, ICCV 2009
http://research.microsoft.com/en-us/um/people/pkohli/papers/lrkt_iccv09.pdf
where (while not the point of the paper) by adding just HOG features to a standard CRF model greatly improves over 4 other recent works.
Anyway, the conclusion to me is that by letting the classifier train over lots of random data, you remove any training bias from your trainingset that your interest point detector might focus on that may not be present in the testset and essentially leave it to the learning algorithm figure out the meaningful and useless features. I think one way to further solidify this hypothesis is to show another axis of classification performance vs object % of the image. Judging from the couple images shown Fig 2, I can buy that a random sampling performs well when the object occupies > 40% of the image, but when it is very small, like for the people in 2nd & 4th image, i am wondering if the interest point detectors work better.
Here is the paper I was talking about in class: http://homes.esat.kuleuven.be/~tuytelaa/Tuytelaars_Schmid-lattice-iccv07.pdf
ReplyDeleteAnother argument for random quantization: projecting your data on random dimensions finds the underlying distribution
ReplyDeletehttp://cseweb.ucsd.edu/~naverma/RPTrees/publication.html
random directions*
ReplyDelete