Streamlining of Viewpoint Synthesis Feature Description Filter Group Using Tensor Decomposition

Keypoint matching is used in a variety of tasks such as a specific object recognition or panoramic image generation. Affine SIFT (ASIFT) generates many affine transformed images from the input image in order to enable affine invariant matching. It describes the scale-invariant feature transform (SIFT) features of the generated image. However, ASIFT must perform multiple costly online computations for affine transformation. In order to solve this problem, the proposed method describes features by convolution of a linear filter. Also, the affine transformation is performed on the linear filter to be used, and the affine feature quantity is described. We use ORB which is luminance based descriptor and GLOH which is gradient based descriptor as linear filter. ORB need to convolve 19,200 times and GLOH need to convolve 20,400 times. In order to reduce the convolution processing, the affine transformed filter is made compact by factorization method. We built a 4-order tensor using the affine transformation filter. The 4-order tensor decomposes into the Tucker decomposition. We reduce dimensions appropriately for each mode. In this way, we propose a saving processing cost and accuracy maintain feature description. From the evaluation experiment, the proposed method reduced the calculation cost by 55.0% compared with singular value decomposition of conventional method.