adaboost features selection from large pool

asked 2013-05-28 03:41:59 -0600

amd_best gravatar image

updated 2013-05-28 03:43:01 -0600

Hi, I want to train adaboost with 10^9 features but it can't stay all to gheter in memory , i procedeed with this approch: for(i=1 to 1000) { -load 10^6 features from the 10^9 pool -select the best 1 features and store it
} but if i test the strong classifier on validation set i have results(correct detection-false positive) worst then this approches: -load 10^6 features from the 10^9 pool -select the best 1000 features and store it

i think the first approches is more exaustive then the second but the second is better on validation set, someone can tell me anything about this anomaly?

edit retag flag offensive close merge delete


you have 10^6 observation of features? How long it is your single feature vector?

yes123 gravatar imageyes123 ( 2013-05-28 06:48:49 -0600 )edit

i want create a ranking of most important features of pedestrian , the pool is very large because i use a features simil to haar, and i can generate a large pool of features, the lenght of vector features descriptor is 10^6 byte

amd_best gravatar imageamd_best ( 2013-05-28 07:52:46 -0600 )edit