TY - GEN
T1 - Boosting the minimum margin
T2 - Digital Image Computing: Techniques and Applications, DICTA 2008
AU - Li, Hanxi
AU - Shen, Chunhua
PY - 2008
Y1 - 2008
N2 - LPBoost seemingly should have better generalization ca- pability than AdaBoost according to the margin theory [12] because LPBoost optimizes the minimum margin directly. Thus far, however, there is no empirical comparison and theoretical explanation of LPBoost against AdaBoost. We have conducted an experimental evaluation on the classi- fication performance of LPBoost and AdaBoost in this pa- per. Our results show that the LPBoost performs worse than AdaBoost in most cases. By considering the margin distri- bution, we present an explanation. Also, our finding indi- cates that besides the minimum margin, which is directly and globally optimized in LPBoost, the margin distribution plays a more important role in terms of the learned strong classifier's classification performance.
AB - LPBoost seemingly should have better generalization ca- pability than AdaBoost according to the margin theory [12] because LPBoost optimizes the minimum margin directly. Thus far, however, there is no empirical comparison and theoretical explanation of LPBoost against AdaBoost. We have conducted an experimental evaluation on the classi- fication performance of LPBoost and AdaBoost in this pa- per. Our results show that the LPBoost performs worse than AdaBoost in most cases. By considering the margin distri- bution, we present an explanation. Also, our finding indi- cates that besides the minimum margin, which is directly and globally optimized in LPBoost, the margin distribution plays a more important role in terms of the learned strong classifier's classification performance.
UR - http://www.scopus.com/inward/record.url?scp=67549150807&partnerID=8YFLogxK
U2 - 10.1109/DICTA.2008.47
DO - 10.1109/DICTA.2008.47
M3 - Conference contribution
SN - 9780769534565
T3 - Proceedings - Digital Image Computing: Techniques and Applications, DICTA 2008
SP - 533
EP - 539
BT - Proceedings - Digital Image Computing
Y2 - 1 December 2008 through 3 December 2008
ER -