This project was the focus of my M.Sc. thesis. The main finding is that higher order image derivatives (aka. image jets) are competitive with traditional local feature descriptors based on gradient orientation histograms (e.g. SIFT). Moreover, image jets allow for a lower dimensional feature descriptor. We evaluate our descriptors on the DTU Robot dataset.
In my MSc thesis I have experimented with other angles to local feature description. Most notably, I have constructed a SIFT-like descriptor using the locally orderless image formulation and the scale-space framework.
Larsen, A. B. L., Darkner, S., Dahl, A. L., & Pedersen, K. S. (2012). Jet-Based Local Image Descriptors. In A. Fitzgibbon, S. Lazebnik, P. Perona, Y. Sato, & C. Schmid (Eds.), Computer Vision – ECCV 2012 (Vol. 7574, pp. 638–650). Springer Berlin Heidelberg. BibTeX / PDF
For evaluating local feature descriptors, I recommend the Matlab framework VLBenchmarks. I have extended it with an interface for the DTU Robot dataset. Beware: There are plenty of evaluative studies of local feature descriptors in the literature. The conclusions of these studies often point in opposite directions. My experience is that the performance of local feature descriptors is very sensitive to the end-application and the data at hand. One should therefore be cautious when drawing conclusions across domains/datasets.