Tensor decomposition of dense SIFT descriptors in object recognition

In machine vision, Scale-invariant feature transform (SIFT) and its variants have been widely used in image classification task. However, the high dimensionality nature of SIFT features, usually in the order of multiple thousands per image, would require careful consideration in place to achieve accurate and timely categorization of objects within images. This paper explores the possibility of processing SIFT features as tensors and uses tensor decomposition techniques on high-order SIFT tensors for dimensionality reduction. The method focuses on both accuracy and efficiency aspects and the validation result with the Caltech 101 dataset confirms the improvement with notable margins.