Directional SIFT -- An Improved Method Using Elliptical Gaussian Pyramid

In this paper, anisotropic scale space is introduced to SIFT method. The method will detect stable elliptical Gaussian blob features of different orientations. Additional feature parameters can be utilized to match features with high probability. New salient features are detected by convolving image with elliptical Gaussian instead circular one. The elliptical Gaussian pyramid is carefully constructed so as to balance elliptical coverage and computational complexity. The new method is tested with both images and videos, which range from indoor objects to outdoor scenes. The results show it can detect several times more salient features than SIFT with same feature quality. This new implementation is not significantly slower than SIFT, and the time complexity is linear with respect to the increased salient features. Multiprocessor or multicore system can be constructed as parallel computing environment, and the method will be as quick as SIFT. SIFT mainly uses descriptors to match, but frame information can also contribute to it. Although new parameters have been introduced, basic neighborhood information will be used to illustrate this idea.