This project started as a spin-off from my M.Sc. thesis in which we experimented with shape index histograms (SIHs) for local feature description. As it turned out, this did not offer any improvements over traditional gradient orientation histograms (from e.g. SIFT). Later, we have found that shape index histograms are good at characterizing texture-like features in images.
I believe that SIHs have the following three selling points compared to alternative texture descriptors:
We participated in a MICCAI 2013 Grand Challenge on mitosis detection. Our method using SIHs came in 2nd after a system using convolutional neural networks. I think this is an excellent result considering the comfortable margin to the other non-deep learning methods.
We entered a method using SIHs in a competition on cell staining pattern classifaction in connection with the 20th IEEE International Conference on Image Processing (ICIP 2013). Our method received the title merit winner as we came in 2nd just 0.1% below the 1st place. Also, our SIHs had superior performance on most classes in the competition.
Larsen, A. B. L., Vestergaard, J. S., & Larsen, R. (2014). HEp-2 Cell Classification Using Shape Index Histograms With Donut-Shaped Spatial Pooling. Medical Imaging, IEEE Transactions on, 33, 1573–1580. PDF BibTeX