Morphological Scale-Invariant Feature Transform
Gert-Jan Ginkel · 2018
This paper proposes a method of substituting the Scale-Invariant Feature Transform descriptors with descriptors based on morphological filters. Multiple configurations are tested for the descriptors, after which the optimal configuration is used to evaluate the new method on the COIL-100, UC Merced Land Use, and Zurich Building Image Database datasets. Not only are the descriptors generated by MorphSIFT less diverse, they also perform worse than the default SIFT descriptors. When using the descriptors in an image retrieval system that creates histograms of size k the MorphSIFT descriptors got the best performance for low values of k, while SIFT descriptors performed better with higher k. This indicates that MorphSIFT descriptors do not have the same encoding power as SIFT descriptors and ultimately have limited application.