[ICCV 2011] Tasting Families of Features for Image Classification英文文献资料.pdf

Tasting Families of Features for Image Classification Charles Dubout and Franc¸ois Fleuret Idiap Research Institute charles.dubout,francois.fleuret@idiap.ch Abstract in 5, vanilla Boosting of stumps over multiple features reaches state-of-the-art performance. However, such tech- Using multiple families of image features is a very ef- niques induce two major practical difficulties: the first is the ficient strategy to improve performance in object detection computational cost of the training, which increases linearly or recognition. However, such a strategy induces multiple with the number of features, and the second is overfitting challenges for machine learning methods, both from a com- the training data. Both are related to the number of features putational and a statistical perspective. which are actually “looked at” during training. The main contribution of this paper is a novel feature We propose here a straight-forward and original strategy sampling procedure dubbed “Tasting” to improve the effi- dubbed “Tasting” to deal with that situation, and use it to ciency of Boosting in such a context. Instead of sampling improve the loss reduction in Boosting. This method sam- features in a uniform manner, Tasting continuously esti- ples a few features from every family before the training mates the expected loss reduction for each family from a per se , and stores their responses over each training sample. limited set of features sampled prior to the learning, and During Boost