Conclusion
In this paper we propose a method to handle rotation in
shape recognition. We effectively adopted a deep learning
framework on the curvature scale space, which computes
the forward and backpropation operations efficiently. We
tested the algorithm on three real datasets, and the performance
on these challenging dataset concludes that modeling
rotation-friendly features facilitates shape recognition.
Acknowledgement: we would like to thank Prof. David
Jacobs and Arijit Biswas from the University of Maryland at
College Park for generously providing the Leafsnap dataset.