Preliminary Local Feature Selection by Support Vector Machine for Bag of Features

This paper proposes a selection method of foreground local features for generic object recognition in “bag of features”. Usually all local features detected from an given image are voted to a histogram of visual words in conventional bag-of-features method. But it may not be good choice because in the standard object recognition task, an image includes target regions and background regions. To distinguish the target from the background, a large number of visual-words are necessary because a variation of local features coming from the background regions is usually large. It is expected that the comparable classification performance will be achieved with a small number of visual words if such unimportant local features can be effectively removed. Although it is difficult to correctly classify all local features into the target and the background, the number of visual-words can be reduced by simply neglecting many of the local features obtained from the background regions which are easily classified by Support Vector Machine (SVM). Experimental results showed the proposed method outperformed the conventional bag-of-features representation with a fewer number of visual-words by neglecting background features by the kernel SVM. The classification performance with linear SVM was also better than the conventional bag-of-feature when the number of visual words was small.