Image Classification with Max-sift Descriptors

In the conventional Bag-of-Features (BoF) model for image classification, handcrafted descriptors such as SIFT are used for local patch description. Since SIFT is not flipping invariant, left-right flipping operation on images might harm the classification accuracy. To deal with, some algorithms augmented the training and testing datasets with flipped image copies. These models produce better classification results, but with the price of increasing time/memory consumptions. In this paper, we present a simple solution that uses Max-SIFT descriptors for image classification. Max-SIFT is a flipping invariant descriptor which is obtained from the maximum of a SIFT descriptor and its flipped copy. With Max-SIFT, more robust classification models could be trained without dataset augmentation. Experimental results reveal the consistent accuracy gain of Max-SIFT over SIFT. The much cheaper computational cost also makes it capable of being applied onto large-scale classification tasks.